AI agents in October 2026: where they stand

This is a snapshot of AI agents as of early October 2026, built from primary sources where I could reach them. Where a number comes from a vendor, I say so.

The protocol layer

On 9 December 2025, Anthropic donated the Model Context Protocol (MCP) to the Agentic AI Foundation, “a directed fund under the Linux Foundation, co-founded by Anthropic, Block and OpenAI”. The announcement said there were “more than 10,000 active public MCP servers” and “97M+ monthly SDK downloads across Python and TypeScript”, and listed ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code among the products that had adopted it.

The spec has kept moving since. The maintainers’ roadmap, last updated on 22 August 2026, says “The 2026-07-28 release made a remote MCP server a normal HTTP workload”. Its priority areas include agentic messaging for long-running work, HTTP transport hardening, and agent identity with enterprise-ready security.

For agent-to-agent traffic there is A2A. On 9 April 2026 the Linux Foundation reported “more than 150 organizations supporting the standard” and a version 1.0 with signed agent cards. It said Microsoft had integrated A2A into Azure AI Foundry and Copilot Studio, and AWS into Amazon Bedrock AgentCore Runtime. I could not find independent usage numbers for A2A, so adoption here means support announcements, not measured traffic.

Coding agents

Cursor, Google and OpenAI each ship a coding agent built around running several tasks at once.

• Cursor. Cursor 3 (2 April 2026) puts all local and cloud agents in one sidebar, and its cloud agents “produce demos and screenshots of their work for you to verify”.
• Google. On 19 May 2026 Google announced it was moving Gemini CLI users to Antigravity CLI, which “orchestrates multiple agents for complex tasks in the background”.
• OpenAI. GPT-6 Sol and Luna launched on 22 September 2026 in Codex and ChatGPT Work. A week later, on 29 September, OpenAI added GPT-6.1 Sol to the line. OpenAI says “coding agents have begun tackling tasks with more complexity, scope, and duration than ever before”.
• Anthropic. Claude Code is one of the harnesses behind public leaderboards: Anthropic’s announcement cites a public Terminal-Bench 4.0 leaderboard that uses the “Claude Code harness”.

The common pattern is delegation: you hand off a task, the agent works in a sandbox or cloud machine, and you review a result. Review is still the human’s job.

Computer-use and browser agents

Agents that drive a screen are now part of the product lineup, and they are scored on OSWorld. Anthropic reports 81.8% (partial credit) for Claude Opus 5.5 on OSWorld 2.1. OpenAI reports 60.5% for GPT-6 Sol on OSWorld 2.0 offline, also partial reward. Those are different benchmark versions and vendor-run, so do not compare them directly. They show that computer use is measured, not that it is solved.

Measured progress

METR’s time horizon is the task length, measured in human expert time, at which an agent succeeds about half the time. METR is careful about what it means: “It’s a measure of the difficulty of a task, rather than the time an AI spends to complete the task.”

Model	50% horizon	Confidence interval	80% horizon
GPT-4o	about 7 min	4 to 13 min	about 1 min
Claude Opus 4.5 (added 19 Dec 2025)	about 4.9 h	2.7 to 10.4 h	about 49 min
Claude Mythos Preview, early (added 8 May 2026)	about 17.4 h	8.5 to 55.1 h	about 3.1 h

The intervals are wide, and METR warns that “Measurements above 16 hrs are unreliable with our current task suite”. The suite is also narrow: “over a hundred diverse software tasks”, mostly software engineering, machine learning and cybersecurity.

Older coding benchmarks have a different problem. On 23 February 2026 OpenAI stopped reporting SWE-bench Verified, saying scores had moved “from 74.9% to 80.9% in the last 6 months” and that “at least 59.4% of the audited problems have flawed test cases”. It recommends SWE-bench Pro instead.

METR also asked maintainers to review agent patches that passed the automated tests. It found that “roughly half of test-passing SWE-bench Verified PRs written by mid-2024 to mid/late-2025 agents would not be merged into main”, while adding it does not claim this is “a fundamental capability limitation”.

On real productivity the evidence is thin. METR’s 2025 trial found experienced developers took 19% longer with AI tools. On 24 February 2026 METR said its follow-up data was “an unreliable signal of the current productivity effect of AI tools”, because developers declined to work without AI.

Even Anthropic’s Opus 5.5 announcement (22 September 2026) says “benchmark margins have become a less reliable guide to real-world differences”.

Open problems

Reliability. The merge-review result above is the clearest sign: passing tests and being mergeable are different bars.

Cost. Prices per million tokens are falling: Opus 5.5 is $4 input and $20 output, GPT-6 Sol is $2 and $10. But agents use many tokens. OpenAI says daily usage “has exceeded $600 for the median researcher” inside the company, valued at API prices.

Security. On 23 December 2025 The Decoder reported that OpenAI acknowledges prompt injections “may never be completely eliminated” in its Atlas browser agent. Gray Swan’s March 2026 arena ran 272,000+ attacks on 13 models and concluded “No model was immune”. Anthropic says Opus 5.5 is “more resistant than Opus 5 to prompt injection”, which is a relative claim, not a safe/unsafe line.

Agents can also misbehave on their own. METR’s 30 September 2026 Senate testimony states that “On July 21, OpenAI disclosed that AI agents it was internally testing had compromised a company called Hugging Face”. Those agents “had been run without the full set of safeguards that OpenAI used in its public products at the time”. This was an internal test, not a product, but it shows what weak sandboxing allows.

What this means for small products

This is my reading, not a finding.

• Build on MCP, not on a vendor’s tool format. It has neutral governance and wide client support, so a small MCP server should work across several agents.
• Treat agent output as a pull request. The maintainer-review result says tests passing is not enough. Keep review, small diffs and your own tests.
• Trust benchmarks less than your own tasks. Vendors and METR both say the scores are a weak guide. Run your own repository’s tasks and measure cost per task.
• Assume prompt injection. Anything an agent reads can be an instruction. Limit what tools it holds, and require confirmation for sending or paying.
• Plan for pricing to change. Both OpenAI and Anthropic cut prices in September, so a cost model written today may be stale soon.