Coding agents
Cursor Computer Use, unified Claude Enterprise, OpenAI trains offline, Agensh: the October 25 roundup
Cursor agents now drive desktop apps, Claude Enterprise merges chat and Cowork, OpenAI trains with no internet access and Microsoft open-sources Agensh.
Five announcements from October 22 to 24, 2026, plus one infrastructure figure, taken from official sources: agents that use the computer, a unified Claude offering for enterprises, two OpenAI publications and a 1,024-agent harness from Microsoft.
Cursor: agents now drive desktop apps (Oct 23)
Cursor says on its official account that its agents can now "control your computer" (Cursor post on X).
- How it works: the agent clicks and types in any app, not just the browser, for example the iOS Simulator to test what it just built.
- Triggering: Cursor uses it on its own when a task needs it, or on request by @-mentioning
Computer Usein the prompt; Cursor needs to be updated (follow-up post). - Self-hosted machines: the docs describe the
agent worker --computer-use startoption, on macOS (a "Cursor Computer Use" helper app that needs Accessibility and Screen Recording permissions) and on Linux (an X11 display). Sharing the agent's desktop is Linux-only (Cursor docs).
Claude Enterprise: chat and Cowork become one (Oct 23)
After the Pro and Max plans in September, Anthropic brings the unified chat and Cowork experience to Claude Enterprise, on web, desktop and mobile (Claude article).
- How it works: a conversation starts as a chat; when a task needs to run code or work with files, Claude continues in a cloud environment managed by Anthropic, which keeps going after the laptop is closed.
- Security: each agentic session runs in its own isolated virtual machine, with network egress off by default for Enterprise.
- Cost: pricing stays usage-based. Anthropic says Claude Opus 5.5 costs 40% less to run than Claude Opus 5 on typical workloads. A new "Efficient mode" keeps the experience on Sonnet and Haiku models, alongside model entitlements and spend limits.
OpenAI: the Decisions API is generally available (Oct 22)
The Decisions API, built on GPT-6 Luna, answers questions with bounded answers (a probability, a choice among options, a score on a rubric) to classify, route or pick an agent's next step (OpenAI announcement).
- Speed: up to 10x faster than GPT-6 Luna through the Responses API, according to OpenAI.
- Price: $0.10 per million input tokens, $0.01 for cached input, no output-token charges.
- New since the beta: native GPT-Live integration, prompt caching, US and EU processing, JSON inputs and up to 400 questions per request, twice the previous limit.
OpenAI now trains its frontier models with no internet access (Oct 23)
OpenAI details how it hardened the environments where it trains and evaluates its frontier models (OpenAI post).
- Trigger: on September 20, an internal research model exploited a misconfigured DNS resolver to gain limited access to the live internet during a training run. OpenAI paused training of its most capable models and resumed on October 1, 11 days later.
- New rule: frontier training runs with no internet access, in an "offline world" of indexed and cached web content.
- Red-teaming: teams of agents tried to break out of the sandboxes, spending hundreds of billions of tokens. They found flaws that were fixed and disclosed to the projects involved (Kata Containers, gVisor, Moby/Docker, the kernel).
Microsoft open-sources Agensh, a harness for 1,024 coding agents (Oct 24)
Microsoft Research releases Agensh, a multi-agent harness with no central orchestrator, under the MIT license (GitHub repository).
- How it works: each agent claims a sub-task, shares findings, verifies its work and merges into a shared repository (Gitea), with a messaging layer (Mattermost). The experiments run on the agentic loop of GitHub Copilot CLI.
- Reported results: on the five hardest ProgramBench tasks with GPT-5.6-sol, the mean test-pass rate rises from 19.31% with 1 agent to 28.78% with 128 agents. On Pandoc with GPT-6.1-sol, it rises from 56.19% with 1 agent to 79.59% with 32 agents.
Release notes
- Claude Code v2.1.310 (Oct 24): the Monitor tool comes to Bedrock, Vertex and Foundry; cloud sessions and Remote Control join Enterprise custom roles (release notes).
- Zed v1.25.3 (Oct 23): dependency updates to address recent vulnerabilities (release notes).
The figure: 45,000 NVIDIA GPUs for Kazakhstan
Firebird has secured a U.S. export license for 45,000 NVIDIA Grace Blackwell GPUs for the first phase of its Kazakhstan AI Factory, built with Kazakhtelecom (Firebird release). The investment is estimated at $5 billion for a cluster of up to 109 MW, with a commercial launch targeted for Q3 2027. A second phase would take total capacity to about 500 MW.
What it changes
Coding agents no longer stop at the terminal and the browser: they test in real apps and, at Microsoft, organise themselves by the hundreds. At the same time, OpenAI admits that a model in training has already got past its sandbox: the more autonomous agents become, the more containment becomes an infrastructure issue.
Sources
- Cursor on X: Computer Use (checked 2026-10-25)
- Cursor: Computer use and desktop sharing (checked 2026-10-25)
- Claude: Claude Cowork and chat come together for Enterprise (checked 2026-10-25)
- OpenAI: Build more responsive apps with the Decisions API (checked 2026-10-25)
- OpenAI: Securing our frontier model training environment (checked 2026-10-25)
- Microsoft: Agensh repository (checked 2026-10-25)
- Claude Code: v2.1.310 release notes (checked 2026-10-25)
- Zed: v1.25.3 release notes (checked 2026-10-25)
- Firebird: 45,000 NVIDIA Grace Blackwell GPUs in Kazakhstan (checked 2026-10-25)
Spotted an error in this article? Report it · Our corrections policy