Research

Agent Platforms and Integration Market: October 2026

No layer owns the enterprise agent control plane in October 2026, and the wire under it has settled. Model, application, integration, and identity vendors each shipped governance for agents they did not build.

What did we find?

  • Disclosed agent revenue sits with application vendors and Microsoft. Salesforce reports Agentforce ARR above $1.5B and ServiceNow targets $1.7B of AI contract value for 2026. Microsoft gave a growth multiple, not a figure. No other model vendor and no integration vendor disclosed an agent revenue line.
  • The surface where people work is going to the model vendors. Workday, ServiceNow, Oracle, and Atlassian shipped their agents into ChatGPT, Claude, Gemini Enterprise, or Copilot rather than asking users to come to them.
  • Governance is the new shelf, and everyone stocked it. Alliances overlap. Okta's Blueprint Alliance includes Salesforce, ServiceNow, AWS, and Google Cloud. ServiceNow's Control Tower Alliance includes Okta, AWS, OpenAI, and Anthropic.
  • Integration vendors are splitting. Enterprise iPaaS leaders sell themselves as the governance layer across other vendors' agents. Workflow tools are becoming tool suppliers inside someone else's agent, priced per tool call.
  • Spend control moved from feature to product. Four vendors shipped hard caps in October, and Salesforce calls part of Agent Fabric a FinOps Dashboard. Buyers now pay for the brake as well as the engine.
  • MCP is the default tool protocol, and its counts do not agree. Directory counts run from 21,617 to 112,900 servers. The protocol went stateless in July, and the NSA said in May that its spread had outpaced its security model.
  • Regulation is a 2027 problem with one 2026 date left. The EU's Digital Omnibus moved Annex III high-risk obligations to 2027-12-02. Colorado replaced its AI Act with a law whose duties start 2027-01-01. The next EU date is 2026-12-02.

This page is the October 2026 edition, kept as it was published on October 30. The current edition is the November 2026 report at the series URL.

The state of the market for tools that businesses use to build, run, govern, and connect AI agents, from no vendor’s side. It covers model vendors, application vendors, integration vendors, frameworks, the protocol layer, and the identity and governance layer. Agentic payments protocols appear only by pointer, because the Agentic Payments Market report covers them. Vendor claims are marked (vendor claim) and are unverified. Third-party evidence is marked (third-party evidence). “Read” means a live page fetch on 2026-10-30 unless another date is given.

What happened in October 2026. This is the first edition, so it opens with what moved in the month.

  • Every layer shipped a control tower in the same month. Google announced an Agent Gateway (10-08). Salesforce added a kill switch and spend limits to Agent Fabric, and SAP launched an Enterprise AI Gateway (both 10-27). ServiceNow formed a Control Tower Alliance and Microsoft counted 50 million registered agents (both 10-28) (section 3).
  • Application vendors put their agents inside model-vendor surfaces. Workday shipped Sana self-service into Gemini Enterprise and Microsoft Copilot, with Claude in early access (10-13). ServiceNow’s Action Fabric now powers a ChatGPT plugin and a Gemini Enterprise integration (section 3).
  • Disclosed agent revenue grew at the suites and at Microsoft. ServiceNow raised its 2026 AI contract-value target to $1.7B. Microsoft reported 120,000 Foundry customers with Foundry revenue more than tripling. Salesforce’s last disclosure stays at $1.5B of Agentforce ARR (section 5).
  • Spend caps became a product line. Google, HubSpot, Anthropic, and Salesforce each shipped hard spend limits or default caps for agent usage inside the month (section 11).
  • Model vendors moved into runtime and into each other’s clouds. OpenAI added computer use to its Agents API beta, and Bedrock Managed Agents powered by OpenAI entered preview on AWS. Anthropic added Managed Agents deployments on AWS. Futurum says Google’s Gemini agent routes across Gemini and Claude models, and Google’s page lists third-party models as early access (section 6).
  • Integration vendors split into platform and supplier. SAP opened Joule Studio to n8n workflows for early adopters, and Zapier bills each MCP tool call as two tasks. Boomi and MuleSoft sell governance over other vendors’ agents (section 6).
  • The protocol layer kept consolidating. MCP and A2A both sit in the Linux Foundation’s Agentic AI Foundation, and A2A shipped a CLI (10-01). SAP opened Joule agents to outside agents through both protocols (section 9).
  • Deal flow stayed with the application vendors. Salesforce agreed to buy Braze for about $5.66B and Rune Technologies, and SAP announced TechWolf (section 12).

1. Summary

No layer owns the enterprise agent control plane in October 2026, and the wire under it has settled. Model, application, integration, and identity vendors each shipped governance for agents they did not build. The same companies joined each other’s alliances. The protocols (MCP for tools, A2A for agents) now sit in one foundation. So connecting an agent to a system is becoming a commodity. The contest has moved to who holds the agent’s identity, its spending limit, and its kill switch.

Seven things define the market in October 2026.

  1. Disclosed agent revenue sits with application vendors and Microsoft. Salesforce reports Agentforce ARR above $1.5B and ServiceNow targets $1.7B of AI contract value for 2026. Microsoft gave a growth multiple, not a figure. No other model vendor and no integration vendor disclosed an agent revenue line (section 5).
  2. The surface where people work is going to the model vendors. Workday, ServiceNow, Oracle, and Atlassian shipped their agents into ChatGPT, Claude, Gemini Enterprise, or Copilot rather than asking users to come to them (section 3).
  3. Governance is the new shelf, and everyone stocked it. Alliances overlap. Okta’s Blueprint Alliance includes Salesforce, ServiceNow, AWS, and Google Cloud. ServiceNow’s Control Tower Alliance includes Okta, AWS, OpenAI, and Anthropic (section 3).
  4. Integration vendors are splitting. Enterprise iPaaS leaders sell themselves as the governance layer across other vendors’ agents. Workflow tools are becoming tool suppliers inside someone else’s agent, priced per tool call (section 6).
  5. Spend control moved from feature to product. Four vendors shipped hard caps in October, and Salesforce calls part of Agent Fabric a FinOps Dashboard. Buyers now pay for the brake as well as the engine (section 11).
  6. MCP is the default tool protocol, and its counts do not agree. Directory counts run from 21,617 to 112,900 servers. The protocol went stateless in July, and the NSA said in May that its spread had outpaced its security model (section 9).
  7. Regulation is a 2027 problem with one 2026 date left. The EU’s Digital Omnibus moved Annex III high-risk obligations to 2027-12-02. Colorado replaced its AI Act with a law whose duties start 2027-01-01. The next EU date is 2026-12-02 (section 10).

2. Categories and terms

Six labels describe the space. Each maps to a different budget owner and shortlist.

LabelWho uses itScope
AI agent platform or agent builderModel vendors, Gartner (“No-Code Agent Builders”)Design and runtime for building, publishing, and managing agents. Gartner defines no-code agent builders as SaaS products with an integrated design and runtime environment, built for citizen developers (Gartner abstract)
iPaaSGartner, Mordor, the integration vendorsCloud integration between applications and data. Now sold with agent builders and MCP gateways attached
Process orchestrationForrester (“Adaptive Process Orchestration”), EverestLong-running workflows that mix people, bots, and agents. Where RPA vendors and iPaaS vendors meet
Agentic AISizing firmsAgent software plus services. Scope varies by firm (section 5)
Agent governance or control towerServiceNow, Salesforce (Agent Fabric), Okta, Microsoft (Agent 365)Discovery, identity, policy, spend, and shutdown for agents across vendors
Agent frameworkLangChain, CrewAI, LlamaIndex, plus model-vendor SDKsCode libraries for building agents, usually sold with a hosted runtime or tracing product

The category decides the buyer. Agent builders reach line-of-business owners and the CIO. iPaaS reaches the integration team. Governance reaches the CISO and the identity team. A vendor that wants the control plane needs all three, which explains why every vendor now claims governance.

Analyst placements, read 2026-10-30. One October evaluation of agent platforms was found, IDC’s MarketScape, and only through a vendor’s announcement. The other placements are adjacent.

ReportDatePlacements
Gartner Magic Quadrant for iPaaS (G00836353) (third-party evidence, through V)2026-03-1618 vendors. Leaders announced by Workato (eighth time), Boomi (12th, “positioned highest for ability to execute” per ITWeb), and Salesforce MuleSoft (tenth year). Jitterbit, Tray.ai, and Celigo announced Visionary (Tray.ai, Celigo, Workato, Jitterbit)
Gartner Emerging Market Quadrant, No-Code Agent Builders (Startup Vendors and Established Vendors editions) (third-party evidence, through V)2026-06-11Glean and Pipefy announced Market Shaper in the startup edition. Boomi and Tray.ai announced Pioneer, and Boomi’s own pages disagree on which edition. Gartner writes that “enterprise giants are the pacesetters” (Gartner article, Boomi, Glean)
IDC MarketScape, Worldwide Agentic AI Platforms 2026 (third-party evidence, through V)2026-10-12Tray.ai announced Major Player (Tray.ai). No other vendor’s placement was read
Forrester Wave, Adaptive Process Orchestration Software, Q3 2026 (third-party evidence, through V)Q3 2026Boomi and UiPath announced Leader (UiPath 2026-10-07, UiPath IR)
Everest Process Orchestration PEAK Matrix (third-party evidence, through V)2026-10-08UiPath announced Leader and Star Performer
Gartner Magic Quadrant, Agentic Software QA Platforms (third-party evidence, through V)2026-10-09UiPath announced Leader. The first Gartner MQ found with “agentic” in its title

The read across these is that analysts rate the integration and orchestration vendors as mature and the agent builders as early. Boomi is a 12-time iPaaS Leader and a Pioneer in agent builders, which is the integration vendor’s position in one line.

3. The stack and where value sits

Buyers assemble seven layers. One vendor rarely owns more than three.

LayerWhat it doesWho sells it (October 2026)
Model and reasoning runtimeRuns the model and the agent loop, hosts tools and sandboxesOpenAI (Agents API), Anthropic (Managed Agents), Google, Microsoft (Foundry), AWS (AgentCore)
SurfaceWhere the person meets the agentChatGPT, Claude, Gemini Enterprise, Microsoft Copilot, Slack, Teams
Agent builder and orchestrationDefines agents, workflows, and hand-offsCopilot Studio, Agentforce, Joule Studio, UiPath, Boomi Agentstudio, Workato, n8n, Zapier, LangGraph, CrewAI
Tools and connectorsLets the agent read and write other systemsMCP servers, iPaaS connector catalogs, application vendors’ own MCP servers
System of recordHolds the data and the permissions the agent acts onSalesforce, ServiceNow, Workday, SAP, Oracle, Atlassian
Identity and policyWho the agent is, what it is allowed to touch, who answers for itOkta, Microsoft Entra Agent ID, ServiceNow AI Passport, SAP Enterprise AI Gateway
Observability, evaluation, and costWhat the agent did, how well, and at what priceLangSmith, Agent Fabric FinOps, AgentCore Evaluations, Anthropic spend analytics, Langfuse (ClickHouse)

The surface is going to the model vendors. Application vendors spent October shipping into model-vendor surfaces. Workday made Sana self-service generally available in Gemini Enterprise, Copilot, Teams, and Slack, with “Sana in Claude Cowork” in early access (Workday, 2026-10-13) (vendor claim).

ServiceNow’s Action Fabric powers a ChatGPT plugin and a Gemini Enterprise integration (ServiceNow Q3 release through SEC) (vendor claim). Oracle’s AI World release (2026-10-26) describes Fusion Companion, which works with Fusion through AI tools such as ChatGPT (Oracle, page blocked, read through a search summary) (vendor claim, through search). An AI Workspace built on ChatGPT plugins is “coming soon” with no date (beri.net, 2026-10-28) (third-party evidence). Atlassian runs an MCP server with “200+ tools” handling “millions of tool calls daily” (Atlassian, 2026-10-06) (vendor claim).

The system of record keeps the permissions, and sells governance on top. Every application vendor that exposed itself to outside agents also shipped a gate in October.

VendorGovernance product (October 2026)Status
ServiceNowAI Control Tower renamed “Super Intelligence Control Tower”. AI Passport verifies “first- and third-party AI agents’ scopes in real-time”. Control Tower Alliance founding members are Anthropic, AWS, Cisco, HiddenLayer, Kong, NVIDIA, Okta, OpenAI, Stripe, WitnessAI, and Zenity (Q3 release, 2026-10-28) (vendor claim)Announced, availability not stated
Salesforce (MuleSoft Agent Fabric)Agent Kill Switch, FinOps Dashboard, Model Wallet with hard spend limits, Cost Optimization Policies, and policies for Azure, Apigee, and Kong gateways (Salesforce, 2026-10-27) (vendor claim)GA. Agent Evals beta November, GA January 2027
SAPSAP Enterprise AI Gateway for authentication, policy, routing, usage, and audit. Third-party agents can call Joule agents through a governed gateway using A2A and MCP (SAP, 2026-10-27) (vendor claim)No dates in the release
WorkdaySigned a definitive agreement to buy C1.ai, an agentic identity platform, price undisclosed. It builds on the “Workday Agent System of Record”, which “already gives companies one place to see and manage their agents” (Workday, 2026-10-13) (vendor claim)Close expected in Workday’s fiscal Q4, which ends 2027-01-31

Model vendors and identity vendors shipped the same thing from the other side. Microsoft reported “over 50 million agents are now in the registry, up over 30 percent quarter over quarter” in Agent 365. Purview has audited 7.6 billion agent activities (Microsoft FY27 Q1, 2026-10-28) (vendor claim). Google’s Gemini agent ships with an Agent Gateway, an Agent Sandbox, and per-project spend caps (9to5Google, 2026-10-08) (vendor claim, through 3P). Okta made Agent SSO generally available on 2026-09-22. It “brings Cross App Access to all SSO customers”, and an Agent Gateway is listed for Q3 (Okta) (vendor claim).

The alliances overlap, which says the plane is federated, not won. Okta’s Blueprint Alliance (2026-09-22) has 12 founding members: Okta, AWS, CrowdStrike, Databricks, Docker, Google Cloud, Lovable, Proofpoint, Salesforce, ServiceNow, Wiz, and Zscaler (Okta) (vendor claim). Five weeks later ServiceNow’s alliance enrolled Okta and AWS too. A vendor in both alliances is hedging. Agent identity looks likeliest to settle with whoever already holds the person’s identity (Okta or Entra) and the object permissions (the system of record). Model vendors then govern what runs on their own surfaces (Speculative, analysis).

4. Buyers and triggers

PersonaBuys onEvidence
CIO or platform engineering leadOne place to build and run agents, model choice, total costMicrosoft counts 120,000 Foundry customers and “over 19,000” on Foundry plus Fabric (section 3 source) (vendor claim)
CISO and identity teamAgent inventory, credentials, kill switch, auditOkta cites Gartner research that “only 13% of organizations think they have the right AI agent governance in place”, linking a Gartner release dated 2026-04-28 (Okta) (third-party evidence, through V). The Gartner page returned 403, so the survey base was not read
Line-of-business owner (service, sales, HR, IT)An agent that closes work inside the system they ownWorkday says “nearly 5,000 customers have activated the Deployment Agent” (Workday, 2026-10-13) (vendor claim)
DeveloperSDK, sandbox, tracing, evaluationLangChain reported over 300 enterprise LangSmith customers in March 2026 (valueaddvc summary) (third-party evidence)
Finance (FinOps)Predictable agent spendSpend caps shipped at four vendors in October (section 11)

Buying triggers, strongest evidence first:

  • Agent sprawl. Microsoft’s registry count (50 million agents, section 3) is the only hard public count of agents under management. Okta cites a Gartner prediction that an average global Fortune 500 enterprise will have “over 150,000 agents in use” by 2028, from the same 2026-04-28 Gartner release (third-party evidence, through V). The Gartner page returned 403.
  • Adoption is broad and shallow. KPMG’s Q3 2026 pulse has 62% of $1B+ US firms building or deploying agents. Mid-market CFOs report 5% agentic tool use (AR Automation Market report). The gap between large-firm pilots and mid-market use is where the builder vendors compete.
  • Cost surprises. HubSpot made spend caps the default on 2026-10-15, with an account limit up to 50% above monthly credits (HubSpot) (vendor claim). A vendor does not ship a default cap unless customers hit the ceiling (Speculative, analysis).
  • Regulatory dates. Colorado’s developer duties start 2027-01-01 and EU Annex III obligations apply from 2027-12-02. Both set 2027 budgets for agents used in employment, credit, and other sensitive areas (section 10).

Three vendors claim most of the Fortune 500. Microsoft says “90% of the Fortune 500 has used Copilot Studio” (vendor claim). Google says “nearly 90% of the Fortune 100 use Gemini Enterprise” (Futurum, 2026-10-09) (vendor claim, through 3P). CrewAI says it is used by “65% of the Fortune 500” (CrewAI pricing page) (vendor claim). All three can be true at once, because “used” counts any team in the company. None is a measure of agents in production.

5. Size and structure

Analysts size AI agents at $7.6 to 11.6B for 2025 to 2026, growing 42 to 50% a year. iPaaS is about $9.2B in 2026, growing about 18%.

Firm (read 2026-10-30)Market labelBaseForecastCAGR
Grand View ResearchAI agents$7.6B (2025)$182.9B (2033)49.6%
Fortune Business InsightsAI agents$8.03B (2025)$251.38B (2034)46.61%
Precedence ResearchAI agents$11.55B (2026)$294.66B (2035)43.57%
Mordor IntelligenceAgentic AI$9.89B (2026, nine months actual plus Q4 projected)$82.27B (2032)42.38%
Mordor IntelligenceiPaaS$9.24B (2026)$20.93B (2031)17.75%
MarketsandMarketsAI agentsPage returned 403

The estimates diverge on scope. Mordor counts agentic software and related services and excludes model APIs, cloud, and conventional RPA. The others do not publish an exclusion list on the open page. Forecast windows run from 2031 to 2035.

Vendor disclosures show the sizings undercount, or count something else. One vendor’s agent line is a fifth of the whole 2025 market as Grand View sizes it.

VendorDisclosed figureSource and caveat
Salesforce”Agentforce ARR exceeded $1.5 billion, up over 240% Y/Y”. Agentforce plus Data 360 ARR “nearly $3.9 billion”. “7.0 billion Agentic Work Units” to dateQ2 FY27, 2026-08-26 (vendor claim). From Q2 the ARR includes Slackbot and Headless 360, so it does not compare with Q1 ($1.2B, through search summary)
ServiceNow”raises AI annual contract value target again for 2026 to $1.7 billion, up 70% from the beginning of the year”. Agentic consumption in production “up more than 13x year-over-year”Q3 2026, 2026-10-28 (vendor claim). A target, not a reported figure
Microsoft”over 120,000 Foundry customers, with revenue more than tripling year over year”FY27 Q1, 2026-10-28 (vendor claim). No dollar figure. No paid Copilot seat update this quarter
Amazon (AWS)AWS sales $46.0B, up 39%. “Tens of thousands of customers currently use AgentCore”, and they deployed twice as many agents in Q3 as in the prior three quarters combinedQ3 2026 release through SEC, 2026-10-29 (vendor claim). No AgentCore revenue figure
Alphabet (Google Cloud)Google Cloud revenue $30.7B, up 103%. First-party model APIs process “nearly 32 billion tokens per minute”Q3 2026 release through SEC, 2026-10-28 (vendor claim). The release does not mention Gemini Enterprise or agents
Workday”AI driving more than 25% of our new ACV”. “More than 5,500 customers now use one or more of Workday’s organic agents, up more than 35% from last quarter” (fiscal Q2 2027)Workday, 2026-08-27 (vendor claim). The “over $100 million” AI figure and an agentic ARR near $600M appear in secondary reports found through search, not in the release (third-party evidence, through search)
UiPathARR $1.938B, up 12%, net new ARR $37MQ2 FY27, 2026-09-03 (vendor claim). Total company, no agent split
LangChainLangSmith ARR estimated at $12 to 16M in June 2025TechCrunch through Fortune, cited by valueaddvc (third-party evidence, estimate, stale)
n8n$40M revenue in 2025Latka (third-party evidence, estimate)

No model vendor separated agent-platform revenue from model revenue in the window. So the visible pool of agent revenue sits with application vendors. Model vendors are paid inside token lines that no public source splits by use.

Concentration. Mordor rates agentic AI concentration “medium” and names Microsoft, Google, and AWS among its leaders. The disclosed revenue puts the measurable share with the large suites and Microsoft. The long tail of builders shows up in valuations, not revenue (section 12).

6. Segments

SegmentWho sells thereWhat decides deals
Model-vendor platformsOpenAI, Anthropic, Google, Microsoft, AWSModel quality, hosted runtime and sandbox, the surface the user already has, cloud commitment drawdown
Application-vendor platformsSalesforce, ServiceNow, Workday, SAP, Oracle, HubSpot, AtlassianData and permissions already in the suite, agents that close work in it, bundled credits
Enterprise iPaaSWorkato, Boomi, MuleSoft (Salesforce), JitterbitConnector depth, governance across vendors, Gartner standing
Workflow automationZapier, Make, n8nSelf-serve price, app count, self-hosting (n8n)
RPA to agentic orchestrationUiPathInstalled bots, long-running process orchestration, QA and testing
Frameworks and developer toolingLangChain (LangGraph, LangSmith), CrewAI, LlamaIndexOpen source adoption, tracing and evaluation, hosted runtime
Identity and securityOkta, Microsoft Entra, Palo Alto Networks (CyberArk)Existing identity footprint, agent discovery, credential brokering

Model vendors moved into runtime, and into each other’s clouds. OpenAI released its Agents API in public beta on 2026-09-10, with OpenAI-hosted or customer sandboxes. At DevDay (2026-09-29) it added computer use and multi-agent support in beta. From 10-09 hosted sandbox compute bills only during active turns (OpenAI changelog) (vendor claim). OpenAI schedules Agent Builder and its Evals platform to shut down on 2026-11-30, pointing users to the Agents SDK or ChatGPT Workspace Agents (OpenAI deprecations) (vendor claim).

“Amazon Bedrock Managed Agents, powered by OpenAI (Preview)” pairs OpenAI models and the Codex harness with AgentCore inside AWS (AWS, Amazon Q3 release) (vendor claim). Anthropic shipped Managed Agents dynamic workflows (10-09) and scheduled deployments on Claude Platform on AWS (10-14) (release notes) (vendor claim). Futurum writes that Gemini “orchestrates across Google’s Gemini models and Anthropic’s Claude models today” (Futurum) (third-party evidence). Google’s own page says “Google Gemini models today, with additional models coming soon” and lists third-party model choice as early access (Google Cloud) (vendor claim, conflict). Microsoft says Codex, Claude Code, and Meta Muse are adopting its new Execution Containers (section 5 source) (vendor claim). Model vendors now sell into each other’s clouds and surfaces, so no model vendor’s platform is a closed stack.

Model vendors priced the runtime separately from the model. AWS AgentCore, Anthropic Managed Agents, and OpenAI containers each carry a runtime unit (section 8). That unit is the start of a platform margin that does not depend on token price (section 11).

Application vendors make their agents callable from outside. Section 3 carries the detail. Salesforce adds its own model, Koa, in pilot, and “Salesforce owns the weights” (Dreamforce recap, 2026-09-28) (vendor claim). Oracle’s Fusion Claw runtime (2026-09-29) uses Gemini and OpenAI models, with more planned, per Oracle’s release read through a search summary (vendor claim, through search). Each suite now chooses which model vendors to admit.

Integration vendors split into platform and supplier.

PatternEvidence (read 2026-10-30)
Governance layer over other vendors’ agentsMuleSoft Agent Fabric scans agents on Agentforce, Bedrock, Vertex, and Copilot Studio (GA January 2026) and added FinOps and a kill switch in October (vendor claim). Boomi’s Agent Control Tower governs Snowflake Cortex agents, and its September release lets Agentstudio agents run on Anthropic, OpenAI, or Gemini models on a self-hosted runtime, early access (Boomi) (vendor claim). Workato’s World of Workato Sydney (10-14) ran under the tagline “The Control and Execution Platform for Enterprise AI: Enterprise MCP + Agentic Orchestration” (vendor claim, through event listing). Celigo launched MCP Management “to Govern AI Assistant and Agent Access Across the Enterprise” on 2026-10-20 (Celigo) (vendor claim). Tray.ai sells an Agent Gateway for MCP and its Merlin Agent Builder as an “Agent Development” add-on (Tray.ai pricing) (vendor claim)
Tool supplier inside someone else’s agentSAP lets early adopters run n8n workflows inside Joule Studio (2026-10-27) after a strategic investment that valued n8n at $5.2B (2026-05-12) (vendor claim). Zapier bills “one MCP tool call = 2 tasks” against the plan’s task pool, with no separate MCP price (Zapier pricing) (vendor claim). UiPath ships an Integration Service MCP Connector (FUSION, 2026-09-23) (vendor claim). Microsoft’s Entra docs name n8n and AWS Bedrock as third-party agent sources to govern (Microsoft Learn) (vendor claim)

The enterprise iPaaS leaders sell governance because their buyer is the integration team, which already owns API policy. The workflow tools sell tool calls because their buyer adopts whichever surface the company picked. MuleSoft’s position is the tell. The strongest integration-vendor governance product belongs to an application vendor (Speculative, analysis).

Frameworks are becoming hosted platforms. In 2026 LangChain launched LangSmith Fleet for building agents without code and made code sandboxes generally available. It also released a managed Deep Agents framework and an LLM gateway in beta (third-party evidence, valueaddvc summary).

CrewAI sells an Enterprise tier priced by execution. n8n released a 3.0 beta on 2026-10-27, a reliability and security release. Chat Hub gives way to “the n8n AI assistant and agents”, and the post reports over 208,000 GitHub stars (n8n) (vendor claim). Make moved its billing unit to credits on 27 August (Make) (vendor claim). LangChain shipped Managed Deep Agents v0.10 with browser use and asynchronous subagents on 10-29 (LangChain blog) (vendor claim).

LlamaIndex moved from framework to document platform. Its October posts cover document parsing and extraction: OpenDocRouter (10-07), the OCRia-1 parsing model (10-22), and a Document Parsing Protocol (10-28) (LlamaIndex blog) (vendor claim). Its price list sells LlamaParse credits, with an agent builder on the Pro and Enterprise plans (LlamaIndex pricing) (vendor claim). It now competes for the document step inside other vendors’ agents, the supplier pattern above (Speculative, analysis).

7. Drivers and restraints

Drivers, ranked by evidence quality.

  1. Application vendors report agent revenue growth. Salesforce’s Agentforce ARR grew over 240% and ServiceNow raised its AI target twice in one year (section 5). These are filed or released figures, the strongest evidence in the market.
  2. Protocol convergence lowers the cost of a connection. MCP and A2A sit in one foundation with every major model vendor as a platinum member (section 9). The Tier 1 MCP SDKs see “close to half-a-billion downloads a month” (MCP blog) (vendor claim).
  3. Consumption growth. ServiceNow’s agentic consumption rose more than 13x year over year. Salesforce counted 3.2 billion Agentic Work Units in Q2, up 97% quarter over quarter (section 5 sources) (vendor claim).
  4. Developer adoption of agent coding. Microsoft says GitHub agents “now open more than a third of all PRs” (section 5 source) (vendor claim). Honeycomb reports “nearly 20% of all monthly interactive queries are now made by agents” (MCP blog, above) (vendor claim).

Restraints.

  • Security of the tool layer. The NSA said in May 2026 that MCP’s spread had outpaced its security model. In 2025 researchers found a 9.6-severity flaw in mcp-remote (CVE-2025-6514) and showed tool-poisoning attacks (Wikipedia summary of NSA, JFrog, and Invariant Labs) (third-party evidence). OWASP published a Top 10 for Agentic Applications on 2025-12-09 (OWASP) (third-party evidence).
  • Cost unpredictability. Spend caps shipped at four vendors in one month (section 11). About half of Salesforce’s Agentforce bookings came from credit replenishment, per call commentary (Nasdaq, through search) (third-party evidence).
  • Governance immaturity. Okta’s 13% figure (section 4) and the overlapping alliances (section 3) show that no shared standard for agent identity exists yet. The MCP roadmap lists agent identity and delegation as open work (section 9).
  • Fragmented counting. Vendors count agents, work units, activities, tool calls, and requests, and none counts the same thing (section 9). Buyers cannot compare platforms on usage.
  • Federal and state conflict in the US. Executive Order 14365 directs a federal task force to challenge state AI laws and names Colorado’s law, since repealed and replaced (section 10). Buyers face rules that can change by litigation.

8. Pricing and deal models

Six pricing models compete. Most vendors mix two or three.

ModelWho uses itExample (read 2026-10-30)
Per seat, agent bundledMicrosoft, Anthropic, Salesforce, AWSAgent 365 $15 per user a month, Microsoft 365 E7 $99 (Microsoft). Claude Enterprise $20 per seat plus API usage, 20-seat minimum (Claude pricing). Agentforce add-ons $125 per user a month (Salesforce)
Credits or unitsSalesforce, Microsoft, Workday, UiPath, Make, HubSpot, Oracle, LlamaIndexSalesforce Flex Credits $500 per 100,000 credits, 20 credits a standard action. Copilot Studio $200 per 25,000 credits a month. LlamaIndex “1,000 credits = $1.25”, Starter $50 and Pro $500 a month (LlamaIndex). Workday Flex Credits unpublished
Per action or per conversationSalesforce$2 per conversation, about $0.10 per standard action
Runtime consumptionAWS, Anthropic, OpenAI, LangChainAgentCore Runtime $0.1276 per vCPU-hour, Managed Agents $0.08 per session-hour plus tokens, LangSmith deployments 0.0675 LSU per vCPU-hour. OpenAI bills hosted sandbox compute only during active turns from 10-09
Per execution or taskn8n, Zapier, CrewAI, Tray.ain8n €20 to €667 a month by executions “regardless of complexity”, Zapier two tasks per MCP tool call. Tray.ai meters integrations, automations, and MCP tools in Tasks and Helix Compute Units, “starts at $50 a month”, with no per-user fee (Tray.ai)
OutcomeHubSpotCustomer Agent and Prospecting Agent moved to outcome-based pricing in 2026 (HubSpot) (vendor claim)

The full row-level price list sits in the Agent Pricing Tracker 2026. Every price above is (vendor claim) from the vendor’s live page unless the tracker says otherwise.

Unpublished prices. Celigo sells Standard, Professional, and Enterprise editions by quote, with Celigo AI, including agent building, in every edition (Celigo) (vendor claim). ServiceNow publishes no list price (third-party evidence). ServiceNow’s 2026 Foundation, Advanced, and Prime tiers include an allocation of assists of unstated size. Contracts sold before those tiers add $25 to $75 per AI-enabled fulfiller a month, a benchmark from observed deals (Redress Compliance, updated 2026-09-24) (third-party evidence). Pro Plus and assist-pack prices were not found at any primary source.

Free runtime is the land grab. SAP made Joule Studio design-time tools free under fair use and custom-agent runtime “free through December 2026 for select customers” (section 3 source) (vendor claim). Salesforce sells Agentforce Foundations at $0 per org.

Google told Futurum it “does not intend to charge a per-seat license when the Gemini agent acts as a coworker in Workspace” (Futurum) (vendor claim, through 3P). Tech-Insider attributes a $30 per user per month figure to Futurum, but the Futurum article contains no $30 figure. Google’s page lists no price, only sales contact or a free 30-day trial of the Plus edition (Google Cloud). The $30 figure is a misattribution and is dropped (corrected 2026-10-30).

Seats bend toward credits. HubSpot lets customers swap seats for credits mid-term (section 4 source). From 2026-12-01, pay-as-you-go becomes the default billing for new Microsoft 365 Copilot Business licenses bought through CSP, moved from 11-02.

The default spending limit is “4,000 Copilot Credits per user per month” (Microsoft Partner Center) (vendor claim). An earlier read dated this to 10-26 through an aggregator (corrected 2026-10-30). Workday’s Flex Credits do not switch agents off when the pool runs out (third-party evidence, beri.net). Seat pricing survives as the entry ticket, and the margin moves to metered use.

9. The agentic shift

Vendor messaging ran from copilot, to named agent, to agent builder, to “control tower” for everyone else’s agents. Proof of agents in production at scale is limited to application vendors and Microsoft. Claim inflation shows in the units.

Salesforce counts Agentic Work Units, Microsoft counts registered agents and audited activities, and ServiceNow counts consumption multiples. Agent Fabric counts “over 1.2 billion requests handled daily”. Atlassian counts “over ten million” human and agent collaborations a month (sources in sections 3 and 5) (vendor claim). None of these units converts to another.

A vendor sits at its highest rung. Rung 2 lists the products still in preview at vendors placed higher.

RungDefinitionWho is there (October 2026)
5. Agent revenue disclosed with production countsAn agent revenue or contract-value figure in a results release, plus production usage countsSalesforce (ARR above $1.5B, 7.0 billion work units, about 2,000 paying production customers added in Q2 per call commentary (third-party evidence)). ServiceNow ($1.7B AI contract-value target, consumption up 13x, no agent customer count)
4. Production usage counts disclosed, no agent revenue lineCustomers, agents, or calls counted in productionMicrosoft (50 million registered agents, 120,000 Foundry customers, revenue multiple only). Google (90% of the Fortune 100, nearly 500 customers past one trillion tokens each, SOMPO with more than 10,000 custom agents (vendor claim, through 3P)). AWS (“tens of thousands” of AgentCore customers, agents deployed in Q3 double the prior three quarters combined, section 5). Workday (more than 5,500 customers on its organic agents, nearly 5,000 on the Deployment Agent, and an “Agent System of Record” it describes as shipped). Atlassian (millions of MCP tool calls daily). MuleSoft Agent Fabric (1.2 billion requests daily). LangChain (300+ enterprise LangSmith customers, March 2026)
3. Generally available and priced, outcomes vendor-claimedShipped product with a price, no production counts disclosedOpenAI (priced hosted sandboxes and AgentKit components, no verified enterprise counts). Anthropic (Managed Agents, and Cowork with chat “generally available for Claude Enterprise plans” on 2026-10-23 (Claude) (vendor claim), no verified enterprise counts). SAP (Joule Studio, “110,000 employees live on Joule Work” inside SAP itself). Oracle (25 Fusion Claw apps in a portfolio of 75 agentic apps, per its 2026-09-29 release through a search summary (vendor claim, through search)). UiPath (Coding Agents GA, no agent counts). HubSpot (outcome-priced agents). Boomi, Workato, Zapier, Make, n8n, CrewAI (65% of the Fortune 500 “used”, a term it does not define). Celigo (agent building in every edition, Celigo Ora GA 2026-09-02, MCP Management 10-20, “thousands of companies” with no agent count) (Celigo) (vendor claim). Tray.ai (Merlin Agent Builder and Agent Gateway for MCP as a quoted add-on, an Apollo story with 40% IT ticket deflection in two weeks) (Tray.ai) (vendor claim). LlamaIndex (agent builder on credit-priced plans, no agent counts). Okta (Agent SSO GA). Palo Alto Networks (Portkey gateway, section 12)
2. Preview, beta, or early adopterNamed product not yet generally availableWorkday Autopilot agents (Talent Acquisition GA early 2027, Payroll and Accounting to early adopters early 2027). Google Gemini agent (private preview). Salesforce Koa (pilot) and Agent Evals (beta November). Anthropic Claude Duet (research preview, 10-21). OpenAI Dots (beta, admin opt-in). OpenAI Agents API (public beta since 09-10) and Bedrock Managed Agents (AWS preview). Microsoft Copilot Studio agent accounts and business process memory (“rolling out now through the Frontier program”, 10-15, Microsoft) (vendor claim). Boomi self-hosted multi-model runtime (early access)
1. Positioning ahead of productCategory claims, alliances, or roadmap with no shipped product behind the claimServiceNow’s Control Tower Alliance and Okta’s Blueprint Alliance (reference architectures and signal sharing, no product). Okta Agent Gateway and Configuration Designer (planned Q3 and Q4 with a no-delivery disclaimer)

Workday’s “Agent System of Record” sits above rung 1. Its C1.ai release of 2026-10-13 names it as a product in use, so the claim is no longer positioning alone.

MCP is the default tool protocol, and its counts conflict.

Measure (read 2026-10-30)ValueSource
Current spec2026-07-28, a stateless core that removes sessions and the initialize handshake, adds server/discover, moves Tasks to an extension, and deprecates Roots, Sampling, Logging, and Dynamic Client Registration in favor of Client ID Metadata DocumentsMCP changelog (third-party evidence, primary)
Next spec dateNone published. The 2026-08-22 roadmap lists agent messaging, transport unification, agent identity and delegation (DPoP, ID-JAG), better primitives, and SDK conformance, and calls Enterprise-Managed Authorization “stable” as an extensionMCP roadmap (third-party evidence, primary)
SDK downloads”close to half-a-billion downloads a month” across Tier 1 SDKs, TypeScript and Python each past 1 billion total. The Register reported 97 million a month in July 2026MCP blog (vendor claim), The Register through Wikipedia (third-party evidence). The two figures conflict, likely on scope
Servers listedOfficial registry 53,203 servers at their latest version, counted by paging the API on 2026-10-30. PulseMCP 21,617. Smithery 27,365+. Glama 112,900. “More than 10,000” in production (The Register, July 2026)MCP registry API (third-party evidence, primary), PulseMCP, Smithery, Glama (third-party evidence)
Vendor call volumeSalesforce 4.5 million MCP calls since launch (May 2026). Atlassian “millions of tool calls daily” (October)The Register through Wikipedia (third-party evidence), Atlassian (vendor claim)
GovernanceMCP donated to the Agentic AI Foundation on 2025-12-09. AAIF hosts MCP, goose, AGENTS.md, A2A, and agentgateway. Platinum members AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, OpenAIAAIF (third-party evidence, primary)

The fivefold spread in server counts shows directories count repositories, forks, and abandoned projects as servers. The usable signal is the vendor call volume and the SDK download trend, not the directory totals.

A2A joined MCP in the same foundation. A2A entered the Agentic AI Foundation as a Growth Stage project on 2026-08-27 and shipped an official CLI on 2026-10-01 (A2A blog) (third-party evidence, primary). Version 1.0 makes the protobuf file normative, renames operations, adds multi-tenancy, and drops implicit and password OAuth flows.

The project tagged v1.0.0 on 2026-03-12 and v1.0.1 on 2026-05-28 (GitHub) (third-party evidence, primary). The technical steering committee has AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP, and ServiceNow (A2A). SAP’s 10-27 gateway is the first October case found of a suite exposing its agents through A2A to outside agents. No count of A2A production deployments was found.

Write actions are the frontier. In receivables, Tesorio, Rillet, and Stripe ship MCP write tools while Tabs and Billtrust stay read-only. Stripe rejects non-agent API keys for MCP after 2026-11-30 (Stripe docs) (vendor claim). The protocol now carries the controls a write needs (issuer validation, credentials bound to one authorization server). Whether a vendor lets an agent write is a product and liability choice, not a protocol gap.

Agent payments protocols (AP2, ACP, Visa and Mastercard agent programs, x402) are covered in agentic-payments-market-2026.md. The AAIF runs an Agentic Commerce working group (AAIF).

10. Regulation

EU AI Act

The Digital Omnibus is law, and it moved the high-risk dates. The Commission proposed it on 2025-11-19, reached political agreement on 2026-05-07, and the amendment entered into force on 2026-07-27 (OJ L 2026/1744) (European Commission, re-read 2026-10-30) (third-party evidence, primary). The page shows the same dates below for general application, Annex III, and Annex I.

DateObligationRead for agent platforms
2025-08-02GPAI model obligations applyModel vendors carry these, platforms inherit the documentation
2026-08-02AI Office and national authorities supervise and enforce. Most remaining provisions applyEnforcement exists now. No agent-specific enforcement found
2026-12-02Article 50(2) machine-readable marking for generative systems placed on the market before 2026-08-02. New prohibitions on non-consensual intimate imagery and CSAM generation (AI Act timeline, updated 2026-08-31)The last 2026 date. Agents that generate content need marking
2027-08-02GPAI models placed on the market before 2025-08-02 must comply. One national sandbox per member stateLegacy models in long-lived agents
2027-12-02Annex III high-risk obligations (biometrics, critical infrastructure, education, employment, migration, credit)Agents that screen candidates or score credit become high-risk systems
2028-08-02Annex I high-risk (AI in regulated products)Mostly outside this market

No AI Office guidance written for agents was found. An agent is regulated by what it does, so a platform that hosts an HR or credit agent inherits Annex III duties through its customer.

United States

ItemStatus (read 2026-10-30)Read
Executive Order 14365, “Ensuring a National Policy Framework for Artificial Intelligence”Signed 2025-12-11. Directs an AI Litigation Task Force to challenge state AI laws, a Commerce list of “onerous” state laws within 90 days, BEAD funding conditions, and FCC and FTC actions. Names Colorado’s law (White House) (third-party evidence, primary)Federal pressure on state rules, with no federal statute behind it
Colorado AI Act (SB24-205), replaced by SB26-189SB25B-004 moved SB24-205 to 2026-06-30 (Colorado General Assembly). That date never applied. SB26-189, “Automated Decision-Making Technology”, signed 2026-05-14, “repeals and reenacts” the 2024 protections. Developer documentation duties start 2027-01-01 (Colorado General Assembly) (third-party evidence, primary). A federal magistrate blocked enforcement of SB24-205 on 2026-04-27 after an xAI suit, per law-firm notes found through search (third-party evidence, through search)Covers technology that influences decisions on education, employment, housing, lending, insurance, health care, and public benefits. A platform whose agents score or recommend in those areas likely counts as a developer (Speculative, analysis)
California CCPA regulations on automated decisionmaking, risk assessments, and cybersecurity auditsApproved 2025-09-22, effective 2026-01-01. Staged compliance dates sit in the regulation text, not read (CPPA) (third-party evidence, primary)Agents that make significant decisions about consumers fall in scope
California SB 53, Transparency in Frontier Artificial Intelligence ActSigned 2025-09-29 (Governor of California) (third-party evidence, primary). Effective 2026-01-01 under California’s default date for statutes. The legislature’s page returned 403, so the bill text was not readApplies to “large frontier developers”, not to platforms
Texas TRAIGA (HB 149)Signed 2025-06-22, effective 2026-01-01 (Texas Legislature Online) (third-party evidence, primary)Government and intent-based duties

Sector and security guidance

ItemDateSource
OWASP Top 10 for Agentic Applications2025-12-09OWASP (third-party evidence)
NIST CAISI posts on agent evaluation transcripts and agent red-teaming, plus an “AI Agent Standards Initiative”2026-02-18 and 2026-03-23NIST CAISI (third-party evidence, primary)
NSA statement that MCP’s spread outpaced its security modelMay 2026Through Wikipedia (third-party evidence, secondary)

Financial-services guidance on agents sits in AR Automation Market report. It includes KPMG’s bank finding that 75% bar agent access to sensitive data without human oversight.

11. Model cost pass-through against platform margin

A platform that resells model calls inside a fixed credit price earns more when model prices fall. A platform that charges for orchestration and lets the customer bring a model key earns the same whatever the model costs. The market runs both models at once.

PositionVendorsWhat the platform charges forWho carries model price risk
Model bundled in creditsSalesforce, Microsoft Copilot Studio, Workday, UiPath, HubSpot, OracleAn action, conversation, credit, or platform unitThe platform. A model price cut widens its margin until it reprices
Bring your own modeln8n, LangSmith, Boomi self-hosted runtime (early access), ZapierExecutions, tasks, traces, or runtime hoursThe customer, through its own key
Model vendor sells the runtimeOpenAI, Anthropic, AWS, GoogleTokens plus a runtime unit (session-hour, container session, vCPU-hour)Shared. The token price falls while the runtime unit holds

Model prices fell inside the window, and the credit prices read did not. OpenAI’s gpt-6.1-terra lists at $1 input and $5 output per million tokens, half of gpt-6.1-sol (OpenAI pricing) (vendor claim). Anthropic cut Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens on 2026-10-07 (section 6 source) (vendor claim). Salesforce’s live price list still shows about $0.10 a standard action and $2 a conversation. If a platform’s credit price holds while the model under it halves, the platform keeps the difference (Speculative, analysis). Salesforce’s claim that Agent Fabric cuts tokens “up to 75%” in test scenarios is that margin offered back to the buyer as a feature (vendor claim).

Spend control is now sold as a product.

VendorControlDate
GooglePer-project spend caps in the Gemini agent2026-10-08
HubSpotSpend caps on by default, up to 50% above monthly credits2026-10-15
Anthropic”Spend limits apply at the organization, group, and member level”, plus an Efficient mode that keeps work on Sonnet and Haiku (Claude) (vendor claim)2026-10-23
AnthropicPooled group budgets, budget alerts at 50%, 75%, 95%, and 100%, and spend by workstream in smart reports (Claude) (vendor claim)2026-10-28
SalesforceModel Wallet hard spend limits, FinOps Dashboard, Cost Optimization Policies2026-10-27
MicrosoftDefault spending limit of 4,000 Copilot Credits per user per month on new Microsoft 365 Copilot Business licenses in CSP (section 8) (vendor claim)Scheduled 2026-12-01
AWSAgentCore Memory moved to GB-based pricing, committed-baseline Runtime pricing “launching by Oct 2026” (AWS)2026-10-06

Four vendors shipped hard caps in one month. The cap moves the overrun risk from the vendor’s invoice to the buyer’s configuration. It also tells the buyer which layer meters the spend. The layer that meters spend is a strong candidate for the control plane (Speculative, analysis).

Runtime units are the platform’s own margin. AgentCore charges per vCPU-hour and per thousand gateway calls, Managed Agents per session-hour, OpenAI per container session (section 8). These units price compute and state, not intelligence, so they hold as tokens get cheaper. A model vendor that wins runtime keeps a margin line its token price war does not touch (Speculative, analysis).

12. Consolidation and funding

Deal timeline (agent platforms, integration, governance). Palo Alto Networks/Protect AI ($500M, July 2025) · ServiceNow/Moveworks (closed December 2025) · Salesforce/Informatica (about $8B, closed 2025-11-18) · ClickHouse/Langfuse (January 2026) · Palo Alto Networks/Chronosphere ($3.35B, closed January 2026) · Palo Alto Networks/CyberArk ($25B, closed February 2026) · UiPath/WorkFusion (2026-02-06) · Palo Alto Networks/Koi Security (about $400M, February 2026) · ServiceNow/Veza (about $1.25B, 2026-03-02) · Palo Alto Networks/Portkey (agreed April 2026, undisclosed) · ServiceNow/Armis (about $7.75B, closed 2026-04-20) · SAP strategic investment in n8n (2026-05-12) · Palo Alto Networks/Console (2026-09-01, IT help-desk agents for Cortex, about $500M reported and not confirmed, SecurityWeek through search) (third-party evidence) · SAP/TechWolf (announced 2026-10-06) · Workday/C1.ai (agreed 2026-10-13, undisclosed, close expected by 2027-01-31) (vendor claim) · Salesforce/Braze (about $5.66B enterprise value, $47.50 a share, 2026-10-20) · Salesforce/Rune Technologies (2026-10-29).

Sources: Workday from its newsroom (vendor claim), Palo Alto deals through Wikipedia (third-party evidence), Langfuse (Langfuse), ServiceNow deals from its 10-Q and IR (vendor claim), Salesforce from its newsroom (vendor claim), SAP from news.sap.com (vendor claim).

What the timeline shows:

  • Application vendors buy the identity and data layers around their agents. ServiceNow bought Veza (identity) and Armis (asset discovery). Workday is buying C1.ai for agent identity. Salesforce bought Informatica (data) and now Braze (engagement). The suites are buying the control-plane parts they did not build.
  • Security vendors buy the agent gateway. Inside fifteen months Palo Alto Networks bought CyberArk (identity), Protect AI (AI security), and Console (agent workflows). It also agreed to buy Portkey, an AI gateway. The gateway between agent and model is becoming a security product.
  • Observability goes to data platforms. ClickHouse bought Langfuse, which says it plans no licensing change for self-hosted users. Tracing data is a database workload, and the database vendors are taking it.
  • Integration vendors take investment from the suites. SAP’s investment in n8n came with a plan to embed it in Joule Studio, the supplier pattern from section 6. No acquisition of an iPaaS vendor was found in the window.

Venture funding concentrates in builders and frameworks. n8n raised $180M at a $2.5B valuation in October 2025. SAP’s investment in May 2026 valued it at $5.2B (Seeking Alpha, Bolsamania) (third-party evidence). Total n8n funding is reported at $240M to $493.5M depending on the tracker. LangChain raised a $125M Series B at $1.25B on 2025-10-20, led by IVP (third-party evidence). Forge Global lists a $3.01B LangChain Series C dated September 2026, with no primary source (section 14). No round for Workato, Boomi, Zapier, or Make was found in the window.

13. Dated checkpoints

DateEventWhy it matters
2026-10-31Okta’s Q3 window closes for Agent Gateway and Shadow AI Agent Discovery GA (quarter basis not stated)Whether the identity vendor ships its own gateway
2026-11-01Microsoft price list preview with SKU IDs for Copilot offers that need no base subscription (vendor claim)First list read of the unbundled Copilot SKUs
November 2026Salesforce Hunter and long-horizon agents GA, Agent Evals and API Evals betaEvaluation becomes part of the application vendor’s governance
2026-11-17 to 11-20Microsoft Ignite (dates from the Microsoft 365 blog banner) (vendor claim)Agent 365 and Copilot Studio agent accounts out of the Frontier program
2026-11-30Stripe rejects non-agent API keys for MCP (Stripe docs). OpenAI shuts down Agent Builder, the Evals platform, and reusable prompt objects (vendor claim)The first hard agent-credential deadline found, and the first model-vendor agent builder retired
2026-11-30 to 12-04AWS re:Invent, Las VegasAgentCore customer counts, Bedrock Managed Agents with OpenAI
2026-12-01Microsoft 365 Copilot offers without a base subscription, priced the same as the add-on offers. Pay-as-you-go default with a 4,000-credit monthly limit on new Copilot Business licenses in CSP (vendor claim)Copilot priced apart from the seat, and the first default cap at Microsoft
2026-12-02EU AI Act Article 50(2) legacy deadline and new prohibitionsThe last EU date of 2026
Early December (date not confirmed)Salesforce Q3 FY27 resultsAgentforce ARR on the new definition, paid production customers
End of December 2026SAP Joule Studio free custom-agent runtime ends for select customersFirst SAP agent runtime price
Late 2026Workday full Sana in Gemini, Copilot, Teams, and Slack. Sana MCP for Copilot early accessWhether a suite’s agent lives mainly in someone else’s surface
2027-01-01Colorado SB26-189 developer documentation duties startA US state duty that can reach agent platforms used in hiring and lending
January 2027Agent Fabric Agent Evals, API Evals, and Custom Agent Scanner GA. Workday’s C1.ai deal expected to close by 2027-01-31MuleSoft’s governance set complete, and Workday’s agent identity layer in house
Q1 2027Workday Sana Concierge GA. Talent Acquisition autopilot GA early 2027Workday’s autopilot agents move from rung 2 to rung 3
2027-08-02EU legacy GPAI models comply, national sandboxes dueLegacy models in agents
2027-12-02EU Annex III high-risk obligations applyHR and credit agents become high-risk systems
Next MCP spec (no date published)Agent identity, server events, Tasks into the coreAgent identity in the protocol, or left to Okta and Entra

14. Evidence limits

  • Web search was limited. Most reads in this edition are direct fetches of known pages, which favors vendor pages over third-party coverage. Absence of an event means “not found in the pages read”, not “did not happen”.
  • Outside-in only. Public research measures what vendors publish, not what they win or what agents do in production.
  • Primary pages blocked fetches. Gartner (including its newsroom), MarketsandMarkets, openai.com, servicenow.com, oracle.com news pages, Microsoft 365 Copilot pricing, and the California legislature returned errors or timed out. A few claims still rest on aggregators or search summaries. Each is labeled and listed below.
  • Counting units do not convert. Agentic Work Units, registered agents, audited activities, requests, tool calls, and consumption multiples measure different things (section 9).
  • Model-vendor revenue is not split. No model vendor discloses agent-platform revenue apart from model revenue, so section 5 cannot size their share.
  • Unverified items.
  • OpenAI. Frontier launch date and pricing, and Connector Registry status (neither appears in the changelog). The Atlassian, Oracle, and ServiceNow deals (headlines only). “1.2 billion weekly users” (Wikipedia shows 900 million weekly in February 2026).
  • Anthropic. Enterprise customer counts. A “900% net revenue retention” claim on one aggregator, excluded as implausible and unsourced.
  • Google. Gemini Enterprise seat prices. Google’s page publishes none, and third-party guides give $21 to about $50 a month by edition, unconfirmed. Gemini 4 Argon GA date (10-13 or 10-16, absent from the Vertex AI release notes read). Whether the Gemini agent routes to Claude today (Futurum against Google’s page, section 6).
  • Microsoft and AWS. Microsoft 365 Copilot list price (the page timed out). AgentCore GA date.
  • Salesforce. The 2,000 paying production customers added in Q2 and the half-of-bookings replenishment share (call commentary, absent from the Q2 release). The Fin and Contentful closes. October GA of Agent Optimizer and Slackforce.
  • ServiceNow, Workday, SAP, Oracle, HubSpot. ServiceNow Pro Plus and assist-pack prices (no primary source, servicenow.com refused connections) and Now Assist contract value. Workday’s $100M AI figure (secondary only), Flex Credits behavior (beri.net only), and the Sana and Paradox closes. SAP’s 400-agent target and AI-unit prices. Oracle’s claims. HubSpot INBOUND 2026.
  • Tray.ai and Celigo. Customer counts and agent prices (both sell by quote).
  • Regulation. California SB 53’s effective date rests on the state’s default date, because the bill page returned 403. The 2026-04-27 Colorado injunction (law-firm notes through search).
  • Figures and counts. The survey base behind Gartner’s 13% and 150,000-agent figures (Gartner page 403).
  • Funding and deals. The LangChain $3.01B Series C listing (no post on LangChain’s blog). LangSmith’s 300 enterprise customers against a separate “6k+ paying customers” figure. Palo Alto Networks’ price for Console.

How was this study done?

Method
Public sources only. Desk study of the tools businesses use to build, run, govern, and connect AI agents. Research window 2026-10-01 to 2026-10-30, with earlier background where the structure needs it. The study is refreshed monthly, and this edition is the first, the October 2026 edition.
Sample
Desk study. 96 public sources: vendor pages, pricing pages and docs, filings and earnings material, analyst and press coverage, and government pages.
Confidence
Every claim comes from a public source and carries a link. Vendor claims are marked (vendor claim) and are unverified. Third-party evidence covers analyst reports, surveys, filings, trade press, and government pages. Web search was limited during the research window, and several analyst, review, and vendor pages blocked automated reads, so some claims rest on press releases, wire copies, or search summaries, and each is labeled. The evidence shows what vendors publish and how analysts and the press describe it. It cannot show what vendors win or what agents do in production. The final section lists the items that stay unverified.

Sources

  1. Gartner abstract
  2. ITWeb
  3. Tray.ai
  4. Celigo
  5. Workato (businesswire.com)
  6. Jitterbit
  7. Gartner article
  8. Boomi
  9. Glean
  10. UiPath IR
  11. Workday, 2026-10-13 (newsroom.workday.com)
  12. ServiceNow Q3 release through SEC
  13. Oracle
  14. beri.net, 2026-10-28
  15. Atlassian, 2026-10-06
  16. Salesforce, 2026-10-27
  17. SAP, 2026-10-27 (news.sap.com)
  18. Workday, 2026-10-13 (newsroom.workday.com)
  19. Microsoft FY27 Q1, 2026-10-28
  20. 9to5Google, 2026-10-08
  21. Okta
  22. Okta
  23. Workday, 2026-10-13 (newsroom.workday.com)
  24. valueaddvc summary
  25. HubSpot
  26. Futurum, 2026-10-09 (futurumgroup.com)
  27. CrewAI pricing page
  28. Grand View Research
  29. Fortune Business Insights
  30. Precedence Research
  31. Mordor Intelligence
  32. Mordor Intelligence
  33. Q2 FY27, 2026-08-26 (investor.salesforce.com)
  34. Q3 2026 release through SEC, 2026-10-29
  35. Q3 2026 release through SEC, 2026-10-28
  36. Workday, 2026-08-27 (investor.workday.com)
  37. Q2 FY27, 2026-09-03 (ir.uipath.com)
  38. Latka (getlatka.com)
  39. OpenAI changelog (developers.openai.com)
  40. OpenAI deprecations (developers.openai.com)
  41. AWS
  42. release notes (support.claude.com)
  43. Google Cloud
  44. Dreamforce recap, 2026-09-28 (salesforce.com)
  45. Boomi
  46. Celigo
  47. Tray.ai pricing
  48. Zapier pricing
  49. Microsoft Learn
  50. n8n (blog.n8n.io)
  51. Make
  52. LangChain blog
  53. LlamaIndex blog
  54. LlamaIndex pricing
  55. MCP blog
  56. Wikipedia summary of NSA, JFrog, and Invariant Labs (en.wikipedia.org)
  57. OWASP (genai.owasp.org)
  58. Microsoft
  59. Claude pricing
  60. Salesforce
  61. Celigo
  62. Redress Compliance, updated 2026-09-24
  63. Microsoft Partner Center (learn.microsoft.com)
  64. Claude
  65. Celigo
  66. Microsoft
  67. MCP changelog (modelcontextprotocol.io)
  68. MCP roadmap (blog.modelcontextprotocol.io)
  69. MCP registry API
  70. PulseMCP
  71. Smithery
  72. Glama
  73. AAIF
  74. A2A blog (a2a-protocol.org)
  75. GitHub
  76. A2A (a2a-protocol.org)
  77. Stripe docs
  78. AAIF
  79. European Commission (digital-strategy.ec.europa.eu)
  80. AI Act timeline (artificialintelligenceact.eu)
  81. White House
  82. Colorado General Assembly (leg.colorado.gov)
  83. Colorado General Assembly (leg.colorado.gov)
  84. CPPA
  85. Governor of California
  86. Texas Legislature Online (capitol.texas.gov)
  87. NIST CAISI
  88. OpenAI pricing (developers.openai.com)
  89. Claude
  90. AWS
  91. SecurityWeek
  92. Wikipedia (en.wikipedia.org)
  93. Langfuse
  94. newsroom (salesforce.com)
  95. Seeking Alpha
  96. Bolsamania

Published .

Cite this

Couch, C. (2026). Agent Platforms and Integration Market: October 2026. Ends at Cash. https://chriscouch.ai/research/agent-platforms-integration-market-2026-october/

BibTeX

@misc{couch2026agentplatformsoct,
  author = {Couch, Chris},
  title = {Agent Platforms and Integration Market: October 2026},
  year = {2026},
  howpublished = {Ends at Cash},
  url = {https://chriscouch.ai/research/agent-platforms-integration-market-2026-october/}
}