Agent Frameworks, Protocols and Coding Agents

Key takeaways

• All facts are as of 12 October 2026. The Model Context Protocol (MCP) is the de facto standard for connecting agents to tools and data, with 42,172 servers in its official registry (entries, not production use). The Linux Foundation's Agentic AI Foundation (AAIF) has governed it since 9 December 2025. Its 2026-07-28 specification removed protocol sessions, so hosted MCP servers built on the old handshake need rework.
• A2A v1.0 (agent-to-agent, 150+ supporting organisations, accepted into the AAIF in August 2026 per its own blog) and AP2 (agent payments, donated to the FIDO Alliance in April 2026) complete the protocol layer. Neither has a named customer deployment in the sources (A2A's own blog names ServiceNow, Salesforce, Atlassian and SAP as platforms using it, a project claim). ACP is in beta under Apache-2.0, latest spec 2026-04-17, maintained by OpenAI and Stripe per its README.
• Choosing a framework now means choosing a runtime and a vendor. LangChain leads on GitHub stars (147.6k), but the four vendor SDKs (Claude Agent SDK, OpenAI Agents SDK, Google ADK, Microsoft Agent Framework) carry MIT or Apache-2.0 code licences and release every week or two. The Claude Agent SDK's README adds that use is governed by Anthropic's Commercial Terms. Anthropic's own guidance prefers simple, composable patterns to heavy frameworks.
• The most urgent migration is OpenAI's. Agent Builder, Evals and reusable prompt objects shut down on 30 November 2026, and Evals become read-only on 31 October. The Assistants API already closed on 26 August 2026.
• Paid coding-agent tiers cluster around $20 a month (GitHub Copilot Pro is $10), and open-source CLIs cost only their model usage. Claude Code itself is proprietary despite its public repository. Codex CLI (98.0M npm downloads a month) and Claude Code (56.3M) lead install volume, and Claude Code's run-rate passed $2.5B in February 2026 (vendor claim). SpaceX's all-stock merger with Cursor, at an implied $60.0B equity value, took effect on 14 August 2026 (SEC filings), and Cognition, maker of Devin and owner of the former Windsurf, is valued at $48B.
• Computer-use agents are not ready for unattended work. In the June 2026 OSWorld 2.0 paper, the best model, Claude Opus 4.8 (max thinking), completed about 20.6% of 108 long-horizon tasks under strict scoring at 500 steps. Later leaderboard rows (Epoch's copy, secondary) put Claude Opus 5 (Max) at 31.4%. Vendor partial-credit scores above 80% use a different version, OSWorld 2.1 (vendor-reported, unverifiable). Prompt injection is unsolved, and OpenAI's security chief has called it "a frontier, unsolved security problem".
• For a business adopting AI automation, the durable assets are MCP servers and evaluation suites that you own. Keep agent loops thin, match frameworks to the cloud you already use, and track deprecation dates. Check licences (Skyvern is AGPL-3.0; Claude Code and the Claude Agent SDK carry Anthropic's Commercial Terms) before you, or a partner on your behalf, host these tools or build them into a product. Prices are in USD. No vendor's own API processes requests in India (OpenAI's India region covers storage at rest only). AWS Bedrock does offer India-only routing (Mumbai and Hyderabad) for the previous Claude generation (Opus 5, Sonnet 5, Haiku 4.5), Kimi K3 and GPT-5.6 Terra/Luna. The newest models (Claude 5.5, GPT-6, Gemini 3.6 and later) have no documented India processing on any cloud read.

Scope, date and evidence labels

This document covers protocols, agent SDKs and frameworks, managed agent runtimes, coding agents, and computer-use and browser agents. Serving, retrieval, memory, observability, workflow tools and voice are covered in Production tool stack, controls and regulation in Governance and security, and model capability in Top models. Stars, versions and downloads were read from the GitHub API, GitHub pages, PyPI and npm on 11 October 2026. They measure developer interest at that date, not production use. Protocol governance, licences, plan prices, adoption figures, benchmarks and India residency were re-checked against primary sources on 12 October 2026. In this document, a vendor claim is a company reporting on its own product. A secondary source is an aggregator, Wikipedia or a news summary. A single-source item has no second confirmation. Where sources conflicted, the most recent primary-source reading is used and the conflict is stated.

The stack has four layers, and lock-in rises with each

Agent software now sits in four layers, and the further up a layer is, the harder it is to leave. At the bottom, protocols such as MCP, A2A, AP2, AGENTS.md, Agent Skills and OpenTelemetry are governed by foundations or published under open licences. Frameworks and SDKs come next. Nearly all are permissively licensed, but model vendors steer the most active ones. Managed runtimes, such as Claude Managed Agents, OpenAI's Agents API and Google's Agent Platform, host the loop, sandbox and memory for a metered fee. At the top, finished agents for coding, browsing and computer use are sold per seat or per credit. An MCP server for your own systems survives a change of model vendor. A workflow drawn in a vendor's visual builder may not survive that vendor's next product decision, as OpenAI's Agent Builder shows.

MCP owns the tool layer, and its July 2026 rewrite forces server rebuilds

Neutral governance and a stateless core

On 9 December 2025 Anthropic donated MCP to the AAIF, a directed fund under the Linux Foundation. Anthropic, Block and OpenAI contributed the AAIF's founding projects: MCP (Anthropic), goose (Block) and AGENTS.md (OpenAI) (Anthropic: donating MCP). Platinum members are AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI (Linux Foundation AAIF release).

Lead maintainers David Soria Parra and Den Delimarsky call the 2026-07-28 specification the largest revision since launch in November 2024 (MCP blog, 2026-07-28). The biggest change is a stateless core: the initialize handshake and the Mcp-Session-Id header are gone, and only an optional server/discover call remains. Streamable HTTP must now send Mcp-Method and Mcp-Name headers. Multi Round-Trip Requests replace server-initiated requests that needed streams held open. List and read responses carry ttlMs and cacheScope so clients can cache them. Authorization adds RFC 9207 issuer validation. Tasks, MCP Apps and Enterprise Managed Authorization move into a formal extensions framework.

Enterprise-Managed Authorization became stable on 18 June 2026. It lets an identity provider grant MCP access, so users connect on first login without a separate OAuth step for each app (MCP blog: enterprise-managed auth). The 22 August roadmap points next to Tasks, subscriptions, webhooks and enterprise readiness (MCP roadmap).

The deprecations start a migration clock. Roots, Sampling and Logging keep at least 12 months of support. The legacy HTTP+SSE transport gets a one-year off-ramp. Dynamic Client Registration gives way to Client ID Metadata Documents. Because servers no longer hold sessions, load balancing is simpler, and ad-tech operators describe rebuilding their servers around the change (PPC Land, secondary).

Adoption is broad but self-reported. The Tier 1 SDKs are TypeScript, Python, Go and C#, with Rust in beta. Together they see "close to half-a-billion downloads a month" (MCP blog, 2026-07-28, vendor claim). At the donation in December 2025, Anthropic counted 10,000+ active public servers and 97M+ monthly downloads across the Python and TypeScript SDKs (Anthropic: donating MCP). The two download figures cover different SDK sets, so they show direction, not exact growth. A full paged pass of the official registry API on 12 October 2026 found 42,172 distinct servers (41,656 with an active latest version, 516 deprecated) and 145,197 version rows. These are registry entries, not production use, and the registry grows daily (MCP registry API).

The Linux Foundation's AAIF release lists Claude, Cursor, Microsoft Copilot, Gemini, VS Code and ChatGPT as MCP clients (Linux Foundation AAIF release; client list seen only as a snippet), and vendor documentation backs this up. OpenAI's Responses API accepts remote MCP servers, and a Secure MCP Tunnel reaches private ones (OpenAI MCP guide). OpenAI has deprecated connector_id for models released after 1 September 2026 (use server_url or tunnel_id; existing models keep connector support; no removal date) (OpenAI API docs). Google offers bring-your-own MCP in Gemini Enterprise and generally available remote MCP servers on Cloud Run (Google Cloud Next '26 wrap-up). Zapier MCP exposes more than 9,000 apps (Zapier MCP).

A2A, AP2 and ACP sit above the tool layer

A2A handles what MCP does not: one agent delegating work to another across vendors. Google created it in April 2025, and the Linux Foundation took over stewardship in June 2025 (CO/AI, secondary). Its technical steering committee includes AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP and ServiceNow (A2A project site). On 27 August 2026 the project's blog said A2A had been "officially accepted as a Growth Stage project" at the AAIF, alongside MCP, goose and AGENTS.md (A2A blog: joining the AAIF, read 11 October 2026). No AAIF page confirming the move was read.

On 9 April 2026 the Linux Foundation called v1.0 the "first stable specification". It added signed Agent Cards and multi-tenancy, and counted "more than 150 organizations" in support, up from more than 50 a year earlier. Named integrations are Azure AI Foundry, Copilot Studio and Amazon Bedrock AgentCore Runtime, with LangGraph and CrewAI cited as interoperable. Production use is described only by vertical, with no named customers (Linux Foundation A2A release). An earlier conflict between sources about the transport is resolved: the specification has a JSON-RPC binding (section 9), a gRPC binding (section 10) and an HTTP+JSON/REST binding (section 11), and it lists 1.0.0 as the latest released version (A2A specification, read 11 October 2026).

Payment protocols are earlier-stage. Google announced AP2 on 16 September 2025 as an extension of A2A and MCP (Google Cloud: AP2). On 28 April 2026 it said it was donating AP2 to the FIDO Alliance, together with Verifiable Intent, which it co-developed with Mastercard (Google: AP2 to FIDO). The project site shows v0.2, and the spec and SDK are Apache-2.0 (AP2 docs). FIDO's own page was blocked, but this document accepts the donation because the donor announced it directly.

ACP, from OpenAI and Stripe, handles checkout. Its repository lists ACP as beta, under Apache-2.0, with a latest spec version of 2026-04-17. The README names OpenAI and Stripe as maintainers, and MAINTAINERS.md also lists Meta (conflict); no foundation is named (ACP repository). ACP reportedly launched in ChatGPT Instant Checkout in February 2026 and moved to an app-based model in March (WeTheFlywheel, secondary). Commentators describe AP2 mandates as verifiable credentials with an x402 crypto extension (Blindpay, secondary). They also see a full purchase as ACP for checkout plus AP2 for proof of authorization, alongside x402, Visa's Trusted Agent Protocol and Mastercard Agent Pay (Crossmint, secondary).

AGENTS.md, Agent Skills and OpenTelemetry package instructions and telemetry

Three newer standards sit beside the wire protocols.

AGENTS.md, contributed by OpenAI, is one of the AAIF's three founding projects (Linux Foundation AAIF release). The project site says AGENTS.md "is now stewarded by the Agentic AI Foundation" and names OpenAI Codex, Amp, Jules, Cursor and Factory among its users. The specification version is not stated (agents.md).

Agent Skills are Anthropic's packages of instructions, metadata and optional scripts that Claude loads only when relevant (Agent Skills overview). They left beta on 19 August 2026, and Managed Agents can load them from a repository's .claude/skills folder (Claude release notes). Agent Skills now has a published specification site (agentskills.io), but its governing body is unidentified. The pattern is spreading anyway: OpenAI's plugins "package skills and MCP configuration" (OpenAI docs index), and Google's Agent Registry indexes "agents, tools and skills" (Google Cloud Next '26 wrap-up). Because a skill can ship scripts, review it like any code dependency.

OpenTelemetry's GenAI semantic conventions name LLM and agent spans in a vendor-neutral way. The GenAI conventions have moved to github.com/open-telemetry/semantic-conventions-genai (the opentelemetry.io page now says "Moved"). They are marked "Development", not stable; semconv v1.44.0 (4 August 2026, Apache-2.0) is the version referenced (OpenTelemetry GenAI conventions; GenAI spans spec).

Frameworks converge on thin SDKs over durable runtimes

Consolidation, cadence and the case for restraint

The framework market has consolidated around a few runtimes.

LangChain 1.0 and LangGraph 1.0 shipped on 22 October 2025. LangChain's create_agent now runs on the LangGraph runtime, legacy chains moved to langchain-classic, and middleware handles human approval, PII scrubbing and retries (LangChain blog, snippet).

Microsoft merged AutoGen and Semantic Kernel into Agent Framework. Agent Framework 1.0 became generally available on 2 April 2026 (python-1.0.0 and dotnet-1.0.0 GitHub releases); the latest releases are python-1.21.0 (8 Oct) and dotnet-1.24.0 (7 Oct) (Agent Framework releases). Secondary sources say Semantic Kernel gets only critical fixes, with support through at least April 2027 (Atlan; is4.ai, secondary). A Microsoft DevBlogs post describes an integration that "shipped ... to v1.0 in C# and Python - fully supported and stable", without a date (Microsoft DevBlogs).

OpenAI is steering code-based builders to its Agents SDK as Agent Builder closes (OpenAI: Agent Builder).

Nine of the frameworks checked released in the week before 11 October 2026: the Claude Agent SDK, CrewAI, Pydantic AI, LangChain, LangGraph, Mastra, Microsoft Agent Framework, Agno and Strands (PyPI 1.59.0 on 8 October). smolagents has not released since 29 May. Letta's last GitHub release is from 14 May, though its PyPI package updated on 9 October.

Version numbers are messy. The Claude Agent SDK is 0.2.165 in Python but 0.3.296 on npm. Google ADK is 2.11.0 on PyPI but 2.2.1 on npm. Microsoft's Python sub-packages still carry 1.0.0b strings next to a 1.21.0 release (Microsoft releases).

Practitioner evidence argues for restraint. Anthropic's engineering guidance says "the most successful implementations use simple, composable patterns rather than complex frameworks". It also notes that its current approach is Managed Agents (Anthropic: Building effective agents). In Stack Overflow's 2025 survey, 84% of developers use or plan to use AI tools, but 52% use no agents or only simpler tools. Among agent builders, Ollama (51%) and LangChain (33%) are the most used (Stack Overflow 2025). LangChain's survey of over 1,300 professionals found 57% with agents in production, and quality was the top blocker, cited by 32% (vendor survey) (LangChain: State of Agent Engineering).

Two first-round figures still circulate and need care. First, the Claude Agent SDK's licence has two layers. Its LICENSE file and PyPI classifier say MIT, but the README's "License and terms" section says use "is governed by Anthropic's Commercial Terms of Service, including when you use it to power products and services that you make available to your own customers" (GitHub: claude-agent-sdk-python, read 11 October 2026). Treat it as MIT code used under Anthropic's commercial terms, not as a plain MIT dependency. Version 0.2.165 (8 October) and 8,249 stars supersede v0.1.75 and about 7.8k stars (GitHub API; PyPI). Second, Google ADK's third-party "2.5.0 on 16 July" figure is superseded by 2.11.0 (2 October) (PyPI: google-adk).

"Not assessed" means the research did not evaluate that framework's capabilities. It is not a negative finding.

Managed runtimes trade control for convenience

Each model vendor now offers to host the agent loop. OpenAI documents three runtimes, rated by integration effort. The Agents API (low effort) runs "an agent with the Codex harness managed by OpenAI". It adds context compaction, multi-agent orchestration, programmatic tool calling, MCP, and a choice of hosted, self-hosted or no sandbox. The Agents SDK (medium effort) runs the loop inside your own application. The Responses API (high effort) is the low-level model interface (OpenAI: Agents overview).

Claude Managed Agents, still in beta, is a "pre-built, configurable agent harness" for long-running and asynchronous work (Claude Managed Agents). Its 2026 additions include per-session spend caps, server-side permission policies and, from 9 October, dynamic workflows in beta, where one agent writes a program that runs many agents (Claude release notes).

Google folded Vertex AI into the Gemini Enterprise Agent Platform at Cloud Next '26 (22–24 April 2026). The platform includes Agent Studio, which exports to ADK, plus Agent Registry, a cryptographic Agent Identity for each agent, and Agent Gateway with Model Armor (Google Cloud Next '26 wrap-up).

Amazon Bedrock AgentCore appears in the sources mainly as an A2A host. Its SDK is current (bedrock-agentcore 1.24.1, Apache-2.0, on PyPI on 7 Oct 2026), but its GA date and prices were not found (PyPI: bedrock-agentcore).

Coding agents settle on $20 tiers, free open-source CLIs and rapid consolidation

Revenue, install volume and trust

Coding agents have the clearest revenue of any agent category. Anthropic launched Claude Code alongside Claude 3.7 Sonnet on 24 February 2025 (Wikipedia: 2025 in AI, secondary) and made it generally available in May 2025 (Anthropic Series G). By 12 February 2026 its run-rate was "over $2.5 billion", more than double the start-of-year figure. Business subscriptions had quadrupled, and enterprises made up over half of Claude Code revenue (Anthropic Series G, vendor claim). The claim that Claude Code wrote 4% of public GitHub commits rests on one secondary source (Second Talent).

OpenAI open-sourced the Codex CLI on 16 April 2025. OpenAI reported more than 5 million weekly active users for Codex on 2 June 2026 (a vendor figure, via Constellation Research; secondary). A figure of about 10 million weekly users on 21 July 2026 combines Codex and ChatGPT Work (secondary). Codex merged into the ChatGPT desktop app on 9 July 2026 (Wikipedia: OpenAI Codex, secondary).

Install volume favours the command-line tools. In the 30 days to 10 October 2026, npm recorded 98.0M downloads of Codex (npm downloads: Codex), 56.3M of Claude Code (npm downloads: Claude Code) and 1.74M of Gemini CLI (npm downloads: Gemini CLI). Automated CI installs inflate these figures, so they are not user counts.

GitHub's Octoverse 2025 says 80% of new developers use Copilot in their first week. It counts more than 1.1 million public repositories using an LLM SDK, including 693,867 created in the past year, up 178% (GitHub Octoverse 2025, vendor claim). Sundar Pichai said about 75% of new code at Google is AI-generated and approved by engineers, up from 25% in 2024 (Semafor; ETV Bharat, vendor claim). Trust lags use: 46% of developers distrust AI accuracy, and 66% say their biggest frustration is answers that are "almost right, but not quite" (Stack Overflow 2025).

Converging prices, shifting owners

Prices have converged on a $20 tier, with credit metering above it; GitHub Copilot's cheapest paid plan is lower, at $10. Claude Code is included in every paid Claude plan, from Pro ($20 a month billed monthly, or $200 a year) to Max ($100 or $200) (Claude pricing; Pro and Team re-read 12 October 2026). Copilot's free tier includes the CLI and agent mode, and agent work draws on "GitHub AI Credits" (GitHub Copilot plans). Aider, Cline and Roo Code are free under Apache-2.0, and their users pay only for model usage (Cline pricing).

Ownership is changing faster than features. SpaceX's merger with Cursor (Anysphere) became effective on 14 August 2026, with 389,289,254 SpaceX Class A shares issued to Cursor holders (SpaceX 8-K, Item 2.01). The April 2026 option agreement in SpaceX's S-1 set an implied equity value of $60.0 billion, with a $1.5B termination fee and an $8.5B deferred services fee under specified conditions (SpaceX S-1/A). The merger agreement followed on 16 June 2026 (8-K). All are SEC filings (primary). Cursor's best-sourced revenue figure is $3B a year in May 2026 (Wikipedia: Anysphere, secondary). A $4B+ figure rests on a single anonymous source (Dealroom). A Reuters headline, seen only in a search index, said OpenAI would cut off its models for SpaceX-owned Cursor (HN index: SpaceX Cursor).

Cognition bought Windsurf in July 2025 and reportedly renamed it Devin Desktop on 2 June 2026 (Wikipedia: Cognition AI, secondary). On 8 September it raised over $2B at a $48B valuation and said its run-rate had grown from $492M in May to almost $900M (Cognition Series E, vendor claim). For a buyer, this means a coding-agent supplier's owner and its access to models can change partway through a contract.

Computer-use agents still fail most long tasks under strict scoring

Computer-use agents work a desktop or browser through screenshots and the accessibility tree, so they need no API. Anthropic's tool became generally available on 19 August 2026 as computer_toolset_20260801, alongside a browser_toolset_20260801 that reads the accessibility tree, forms and tabs. Older tool versions return HTTP 400 errors on Opus 5.5 and Haiku 5.5, so integrations break at model upgrades (Claude release notes). OpenAI keeps its computer tool, but for GPT-6 Astra it recommends Playwright or PyAutoGUI code execution instead (OpenAI: Computer use). Gemini computer use reached public preview in Gemini 3.5 Flash on 24 June 2026 (Gemini API changelog).

Benchmarks depend on the metric

The OSWorld 2.0 paper (arXiv 2606.29537, June 2026) set 108 long-horizon workflows (median about 1.6 human hours and about 318 tool calls each) and scored strict, all-or-nothing completion at 500 steps. The best model, Claude Opus 4.8 (max thinking), finished about 20.6% (54.8% partial); GPT-5.5 plateaued near 13% (OSWorld 2.0 paper). Later leaderboard rows (Epoch's copy, secondary) show Claude Opus 5 (Max) at 31.4% strict (68.3% partial) and GPT-5.6 Sol (Max) at 27.3% (62.7%). There are no Opus 5.5, Fable 5.1, Sonnet 5.5 or GPT-6 rows (Epoch benchmark data). Vendor scores are far higher because they use partial credit, other setups and newer models. Anthropic also reports that Fable 5.1 scored zero on OSWorld 2.0 tasks where its safeguards intervened (Claude Fable). Safety layers can lower automation success rates, not just benchmark scores.

Prompt injection is the harder limit

Anthropic's documentation warns that "Claude will follow commands found in content even when they conflict with your instructions". It advises isolating Claude from sensitive data and actions, while claiming trained resistance and an extra layer of defence (vendor claim) (Anthropic computer-use tool). Public research on AI browsers backs the warning. In August 2025 Brave showed that Perplexity Comet passed webpage text to its model mixed in with the user's instructions, so hidden text on a page could reach the user's email open in another tab (Brave: Comet; the post shows no date). In October 2025 Brave showed nearly invisible text in screenshots working as commands, and reported an Opera Neon injection that it withheld at Opera's request (Brave: screenshot injections). At the October 2025 launch of ChatGPT Atlas, OpenAI's chief information security officer called prompt injection "a frontier, unsolved security problem" (Simon Willison: OpenAI CISO on Atlas). One secondary source says Atlas was exploited soon after launch (Delante, secondary), but no specific exploit write-up was verified. Atlas shut down on 9 August 2026, after a retirement announced on 9 July 2026, with its features moved into a combined ChatGPT desktop app (Wikipedia: ChatGPT Atlas citing Lifehacker and TechRepublic; secondary; no OpenAI page read).

Two widely repeated claims could not be confirmed. OpenAI's reported view that browser injection "may never be fully patched" was not confirmed at the source (softwareseni, secondary). A claimed Trail of Bits audit of Comet is absent from that firm's blog feed (Trail of Bits feed).

Eleven incidents show where agent controls belong

The pattern repeats across incidents. An agent with broad tool access reads untrusted text, then acts with the user's credentials.

The controls follow from that pattern. OWASP's LLM06:2025 entry on excessive agency calls for narrow tools instead of open shells, read-only identities where possible, and OAuth scopes no wider than the task needs. It also calls for human approval of high-impact actions, authorization enforced in downstream systems, and rate limits (OWASP LLM06:2025). OWASP also added a Top 10 for Agentic Applications on 9 December 2025 (OWASP Agentic Top 10). Simon Willison advises running unattended coding agents in a container, VM or OS sandbox with restricted network egress, keeping SSH keys and cloud credentials out of reach (Simon Willison). Invariant's free MCP-Scan (Apache-2.0, now maintained by Snyk) checks servers for poisoned tool descriptions (Snyk). No malicious MCP server has been verified in the wild. See Governance and security.

Deprecations and ownership changes are the practical lock-in risk

OpenAI's shutdowns come first. On 3 June 2026 OpenAI announced the end of Agent Builder and Evals. Evals become read-only on 31 October. Agent Builder, the Evals dashboard and API, and v1/prompts shut down on 30 November 2026 (OpenAI deprecations). The same page names Promptfoo as the migration path for Evals. You can export a workflow as Agents SDK code or rebuild it as a ChatGPT Workspace Agent. OpenAI warns that export "does not convert your workflow graph or guarantee that every behavior transfers unchanged" (OpenAI: Migrate from Agent Builder). The Assistants API had already closed on 26 August (OpenAI changelog). As of 12 October, teams still on Agent Builder have about seven weeks.

Model churn comes second. Anthropic shipped about one model tier a month from June to October 2026, and each came with breaking changes. On Haiku 5.5, for example, temperature, top_p, top_k, manual budget_tokens and prefill "can each return a 400 error" (Claude release notes). Access itself can vanish. A US export-control directive made Anthropic switch off Fable 5 and Mythos 5 for every customer from 12 June to 1 July 2026 (Anthropic: Fable and Mythos access). Google limits Gemini 2.5 to recent users and retires agent builds and previews on fixed dates (Gemini API changelog).

Licences and ownership come third. Projects can stop: the owner of Flowise, a low-code agent builder with 55.5k stars, archived it on 13 August 2026; it reached end of life on 31 August 2026, and its npm packages and Docker images are deprecated (Flowise discussion). Licences can restrict who may host a tool for whom. Skyvern is AGPL-3.0, which carries obligations when it is offered as a hosted service. n8n's licence allows use only "for your own internal business purposes or for non-commercial or personal use", and distribution or provision to others only "free of charge for non-commercial purposes", so a provider that runs n8n for you as a paid managed service needs a commercial agreement with n8n (inference from the text) (n8n licence). Some licences are easy to misread. GitHub reports NOASSERTION for Mastra and the Vercel AI SDK, but their LICENSE files say Apache-2.0, and Mastra's ee/ folders fall under the separate Mastra Enterprise Edition License v2.0 (effective 22 Sep 2026). Claude Code's npm licence field points to its README (npm: Claude Code), and its public repository is not open source: the LICENSE.md file reads "All rights reserved" and puts use under Anthropic's Commercial Terms (Claude Code LICENSE.md, read 11 October 2026).

Compatibility shims carry their own traps. DeepSeek's Anthropic-compatible endpoint lets Claude Code switch providers with a base-URL change. But it silently maps claude-opus names to deepseek-v4-pro, ignores cache_control, does not support MCP blocks, and sends data to infrastructure hosted in China (DeepSeek Anthropic API guide; see DeepSeek). xAI has fully deprecated its own Anthropic compatibility layer (xAI docs). A shim is a convenience, not a portability strategy.

Decision guide: if you are building X, start with Y

What this means for your business

First steps and deadlines

Move off OpenAI's closing products. Evals become read-only on 31 October 2026; Agent Builder, Evals and v1/prompts shut down on 30 November. Export does not guarantee unchanged behaviour, so retest each migrated workflow.

Build connectors as MCP servers. An MCP server for your CRM, ERP or document store works from Claude, ChatGPT, Gemini, Copilot and Cursor and survives a change of model vendor, which a vendor's visual workflow may not. Build new servers on the stateless 2026-07-28 spec; hosted servers on the old handshake need rework.

Decisions to make

Match the framework to your cloud: Microsoft Agent Framework for Azure, ADK for Google Cloud, the Agents SDK for OpenAI-only work, and LangGraph (MIT, durable runtime) for neutral or multi-model work. A single-purpose assistant may need no framework.

Keep a second vendor's model ready. Monthly releases with breaking changes, and a 19-day government-ordered withdrawal of two frontier models, make a fallback necessary. Keep regression tests for both models in your own repository, not on a vendor dashboard, as OpenAI's Evals shutdown shows. See Costs and Top models.

What to budget

Coding-agent seats cluster around $20 a month; open-source CLIs cost only model usage. Managed runtimes add metered fees on top of tokens. Gemini 3.6, 3.7 and 3.8 Flash prices double on 1 January 2027. Most SDKs are pre-1.0, so budget for regular upgrades.

Risks and what to test

Security. Require least-privilege tools with one identity per agent, human approval for every write, payment or external send, sandboxed runtimes with an egress allowlist, no "trust all tools" flags, and an MCP-Scan check on every server. Apply the same rules to developer laptops and CI pipelines, where the s1ngularity attack ran. Supervise browser and computer-use agents while strict OSWorld 2.0 success is 20–31%. See Governance and security and the Playbook.

Licences. Check AGPL, n8n's internal-use clause, Mastra's ee/ licence, Claude Code's proprietary licence and the Claude Agent SDK's commercial terms before you, or a partner on your behalf, host any of them.

India. Prices are in USD. Anthropic reportedly began rupee billing for Claude consumer and Team plans on 13 Jul 2026, with GST included (Pro ₹2,399 a month; Team Standard ₹2,399 a seat a month billed annually, ₹2,999 monthly; single-source, Techstory); INR invoicing for the API is unresolved. GST on foreign AI APIs was not read from any statute: the usual position is reverse charge for B2B imports of services, so get tax advice. On the Claude API, inference_geo offers only global or US processing, with US at 1.1x the price (Claude data residency). Anthropic's compliance page (read 11 Oct 2026; not re-readable 12 Oct) lists India among its Asia-Pacific residency regions via cloud partners, but names no India region ID (Claude regional compliance). On AWS Bedrock, India-geography cross-region profiles (in. prefix) route only to Mumbai (ap-south-1) and Hyderabad (ap-south-2) for Claude Opus 5, Sonnet 5 and Haiku 4.5, and also for Kimi K3 and GPT-5.6 Terra/Luna. The Claude 5.5 models, Fable and Mythos are global-only from the India regions (Bedrock regional availability; Bedrock model card: Claude Sonnet 5). AWS's regional-availability page lists Claude Opus 5 and Sonnet 5 as In-Region in Mumbai on the bedrock-mantle endpoint, but the model cards omit it, so confirm in the console. AWS says to call GetInferenceProfile rather than trust the prefix, and that prompts and outputs may be stored in opt-in Regions for abuse detection (Hyderabad is opt-in, so still India). Google Cloud documents India processing for no Claude model (Google Cloud data residency), and Azure Foundry documents only a US Data Zone for Claude; its only India region, South India, offers GPT-5.6 and GPT-6 only as Global deployments (Foundry deployment types; Foundry region availability). From 13 May 2027, DPDP Rules 6(1)(e) and 8(3) require logs to be kept for one year (DPDP Rules 2025), which reaches agent and MCP-server logs that hold personal data (inference). OpenAI's regional processing costs 10% more and is offered only in the US, the EU and the UAE. Its India region (in.api.openai.com) covers storage at rest only and needs OpenAI's approval (OpenAI data controls). OpenAI's capacity deal with Tata's HyperVault does not by itself keep API traffic in India (TechCrunch). In practice, in-country processing today means Bedrock's in. profiles and the previous model generation; get residency terms in writing before you rely on them.

Who can build it

Large integrators and model vendors also deploy agents. Anthropic says Cognizant is "deploying Claude to 350,000 employees" (Anthropic: Bengaluru, vendor claim). TCS will use Codex (TechCrunch) and reported annualised AI revenue above $3B for Q2 FY27 (8 Oct 2026; secondary, several outlets agree; Outlook Business). OpenAI launched DeployCo, a deployment and integration company, in May 2026 (Wikipedia: OpenAI, secondary). For in-house teams, India added more than 5 million GitHub developers in 2025 (GitHub Octoverse 2025). See India and Companies.

Questions to ask any vendor or implementation partner

• Does your business own the MCP servers and evaluation suite, and are they built on the 2026-07-28 spec?
• Can each agent switch to a second vendor's model, with regression tests for both?
• Which licences cover each component, and who holds any agreement needed to host it?
• Which actions need human approval, and how are agents sandboxed?
• Where are your data and logs processed and kept, and on which model generation?

Open questions and unverified items

• AP2 and ACP. FIDO's own page on the AP2 donation was not read; Google, the donor, confirms it. ACP is in beta, spec 2026-04-17, Apache-2.0; its README names OpenAI and Stripe as maintainers, but MAINTAINERS.md adds Meta. AGNTCY is "a Series of LF Projects, LLC", but its version and any link to the AAIF are unverified.
• A2A. The move to the AAIF still rests only on the project's blog (single-source). The blog names ServiceNow, Salesforce, Atlassian and SAP as platforms using A2A, plus native support on Google Cloud, Bedrock AgentCore and Azure AI Foundry, but no customer deployment.
• AAIF membership. Only the Platinum tier was read; the wider member list is unverified.
• Newer standards. AGENTS.md is stewarded by the AAIF, but its specification version is not stated. Agent Skills has a specification site (agentskills.io) but no identified governing body.
• Licences. How Anthropic's Commercial Terms interact with the Claude Agent SDK's MIT LICENSE file needs a lawyer's reading before you build the SDK into a product or service for your own customers.
• Pricing. Codex/ChatGPT plan pages and Windsurf are still unreadable. Copilot Business and Enterprise seat prices are not isolated. Jules shows task limits (15 free, 100 Pro) but no prices. Antigravity's AI Pro and Ultra prices are missing ($30 seats apply to Gemini Enterprise customers). Lovable shows only its credit structure. Per-session computer-use costs are missing.
• Adoption. Codex has 5M+ weekly active users (OpenAI report of 2 Jun 2026 via Constellation Research; secondary, vendor figure; OpenAI's own report not read); a combined Codex plus ChatGPT Work figure is about 10M (21 Jul 2026). Claude Code user counts are unpublished. Cursor's $3B versus $4B+ conflict is unresolved, and the 4% of GitHub commits claim is single-source.
• Cursor and Grok Build. The OpenAI–Cursor cut-off headline was not opened. The Grok Build repository-upload report rests on a single-source chain.
• Benchmarks. The OSWorld 2.0 paper figures are read from the arXiv abstract (primary). Opus 5's 31.4% comes only from Epoch's copy of the leaderboard (secondary). No independent OSWorld 2.1 or OSWorld-Verified board was read, so vendor OSWorld 2.1 figures are unverifiable.
• AI browsers. Comet: no discontinuation found (secondary; Perplexity's page is blocked). The Atlas shutdown (9 Aug 2026) rests on Wikipedia citing Lifehacker and TechRepublic; no OpenAI page was read. Brave's Comet post shows no date and no "not fully mitigated" update. OpenAI's "never fully patched" quote, the Trail of Bits audit and the University of Washington same-origin study remain unverified.
• Framework depth. Per-framework strengths, weaknesses and production users were not independently researched.
• Other platforms. AgentCore's GA date and prices are unresolved (the pricing page is JS-rendered; bedrock-agentcore 1.24.1 is on PyPI). The Strands feature set and DeepSeek's agent-harness preview are still not researched.
• OpenAI MCP connectors. The connector_id deprecation has no removal date.
• India. INR invoicing for the Claude and OpenAI APIs (rupee prices for Claude consumer and Team plans are single-source) and GST treatment remain open (Bedrock routes the previous Claude generation, Opus 5, Sonnet 5 and Haiku 4.5, inside India; no cloud documents India processing for the Claude 5.5 models).
• Bedrock documentation conflict. AWS's regional-availability page lists Claude Opus 5 and Sonnet 5 as In-Region in Mumbai on bedrock-mantle, but both model cards' bedrock-mantle tables omit Mumbai. Mumbai prices are not captured.
• Google Cloud conflict. The data-residency page lists Singapore ML processing for Claude Opus 5.5, Sonnet 5.5, Sonnet 5, Opus 5 and Fable 5.x, but the locations page does not.

Sources

1. Anthropic: donating MCP
2. Linux Foundation AAIF release
3. MCP blog, 2026-07-28
4. MCP blog: enterprise-managed auth
5. MCP roadmap
6. PPC Land
7. MCP registry API
8. OpenAI MCP guide
9. OpenAI API docs (full text)
10. Google Cloud Next '26 wrap-up
11. Zapier MCP
12. CO/AI
13. A2A project site
14. A2A blog: A New Chapter for A2A, joining the Agentic AI Foundation (27 Aug 2026)
15. Linux Foundation A2A release
16. A2A specification
17. Google Cloud: AP2
18. Google: AP2 to FIDO
19. AP2 docs
20. ACP repository
21. WeTheFlywheel
22. Blindpay
23. Crossmint
24. AGENTS.md project site
25. Agent Skills overview
26. Claude release notes
27. agentskills.io
28. OpenAI docs index
29. OpenTelemetry GenAI conventions
30. OTel GenAI spans spec
31. AP2 repo
32. LangChain blog
33. Agent Framework releases
34. Atlan
35. is4.ai
36. Microsoft DevBlogs
37. OpenAI: Agent Builder
38. Microsoft releases
39. Anthropic: Building effective agents
40. Stack Overflow 2025
41. LangChain: State of Agent Engineering
42. GitHub: anthropics/claude-agent-sdk-python (README, "License and terms")
43. GitHub API: anthropics/claude-agent-sdk-python
44. PyPI: claude-agent-sdk
45. PyPI: google-adk
46. Claude Agent SDK overview
47. GitHub API: openai/openai-agents-python
48. GitHub API: google/adk-python
49. GitHub API: microsoft/agent-framework
50. PyPI: strands-agents
51. GitHub API: langchain-ai/langchain
52. GitHub API: langchain-ai/langgraph
53. GitHub API: crewAIInc/crewAI
54. GitHub API: run-llama/llama_index
55. GitHub API: agno-agi/agno latest release
56. GitHub API: stanfordnlp/dspy
57. GitHub API: huggingface/smolagents
58. GitHub API: deepset-ai/haystack
59. GitHub API: letta-ai/letta
60. GitHub API: pydantic/pydantic-ai
61. npm downloads: Vercel AI SDK (ai)
62. Vercel AI SDK LICENSE
63. npm downloads: @mastra/core
64. Mastra LICENSE.md
65. Mastra ee/LICENSE
66. OpenAI: Agents overview
67. Claude Managed Agents
68. Claude pricing
69. Claude data residency
70. OpenAI Agents API
71. OpenAI pricing
72. OpenAI data controls (data retention and residency regions)
73. Agent Platform pricing
74. Microsoft Learn: Copilot Studio billing and licensing
75. xAI Grok Bot
76. PyPI: bedrock-agentcore
77. Wikipedia: 2025 in AI
78. Anthropic Series G
79. Second Talent
80. Constellation Research: Codex users
81. Wikipedia: OpenAI Codex
82. npm downloads: Codex
83. npm downloads: Claude Code
84. npm downloads: Gemini CLI
85. GitHub Octoverse 2025
86. Semafor
87. ETV Bharat
88. GitHub Copilot plans
89. Cline pricing
90. SpaceX 8-K (14 Aug 2026)
91. SpaceX S-1/A
92. Wikipedia: Anysphere
93. Dealroom
94. HN index: SpaceX Cursor
95. Wikipedia: Cognition AI
96. Cognition Series E
97. Claude Code LICENSE.md
98. GitHub: anthropics/claude-code
99. OpenAI deprecations
100. GitHub: openai/codex
101. Gemini CLI: quota and pricing
102. GitHub: google-gemini/gemini-cli
103. Cursor pricing
104. Devin pricing
105. Google Antigravity pricing
106. Google Jules
107. Kiro pricing
108. Factory pricing
109. Amp pricing
110. Roo Code / Roomote pricing
111. GitHub: cline/cline
112. GitHub: Aider-AI/aider
113. GitHub: RooCodeInc/Roo-Code
114. Wikipedia: Grok
115. Replit pricing
116. Bolt pricing
117. v0 pricing
118. Wikipedia: Lovable
119. TechCrunch: Replit's Amjad Masad interview (1 May 2026)
120. OpenAI: Computer use
121. Gemini API changelog
122. arXiv 2606.29537: OSWorld 2.0 paper (abstract)
123. Epoch benchmark data
124. Claude Fable
125. BenchmarkList
126. Anthropic Opus 5.5 page
127. Gemini 3.8 Flash card
128. Constellation Research
129. Snorkel
130. Anthropic computer-use tool
131. Brave: Comet
132. Brave: screenshot injections
133. Simon Willison: OpenAI CISO on Atlas (22 Oct 2025)
134. Delante
135. Wikipedia: ChatGPT Atlas
136. softwareseni
137. Trail of Bits feed
138. Claude in Chrome
139. Google: Argon
140. GitHub: browser-use/browser-use
141. PyPI: browser-use
142. Browser Use pricing
143. GitHub: browserbase/stagehand
144. npm downloads: Stagehand
145. GitHub: microsoft/playwright-mcp
146. npm registry: @playwright/mcp
147. npm downloads: Playwright MCP
148. GitHub: Skyvern-AI/skyvern
149. Skyvern pricing
150. Opera Neon
151. Invariant Labs: MCP tool poisoning (1 Apr 2025)
152. Invariant Labs: GitHub MCP exploit (26 May 2025)
153. The Register
154. Wiz: malware invoking AI
155. BleepingComputer
156. Cloudflare: response to Salesloft Drift incident
157. arXiv 2509.10540: EchoLeak
158. Anthropic: disrupting AI-orchestrated espionage (13 Nov 2025)
159. ABC News
160. Simon Willison
161. OWASP LLM06:2025
162. OWASP Agentic Top 10
163. Snyk: acquisition of Invariant Labs
164. OpenAI: Migrate from Agent Builder
165. OpenAI changelog
166. Anthropic: Fable and Mythos access
167. Flowise discussion
168. n8n licence
169. npm: Claude Code
170. DeepSeek Anthropic API guide
171. xAI docs
172. Google pricing
173. Claude data retention
174. Techstory: Claude rupee billing
175. Claude regional compliance
176. AWS Bedrock: model support by Region
177. AWS Bedrock model card: Claude Sonnet 5
178. Google Cloud: Agent Platform data residency
179. Microsoft Learn: Foundry deployment types
180. Microsoft Learn: Foundry models by region
181. DPDP Rules 2025 (gazette)
182. TechCrunch: OpenAI and Tata HyperVault (18 Feb 2026)
183. Anthropic: Bengaluru
184. Outlook Business: TCS Q2 FY27 results
185. Wikipedia: OpenAI