Cloud agents, plugin packages, and a model migration deadline

Agents that manage other agents

The week's biggest shift came from Cursor, which introduced Projects, a left-hand-nav surface where a coordinator agent plans work, delegates it to parallel subagents on cloud machines, and returns finished work for review. Rather than editing code itself, the coordinator holds shared context files that sync between cloud and local machines and can act on Slack channels, schedules, or pull requests without a prompt. It's a meaningful repositioning: the unit of work stops being a chat turn and becomes a body of work that persists over months.

Cognition moved in the same direction with smaller pieces. Devin now bundles skills, rules, hooks, and MCP servers into installable plugin packages, and archiving a session closes the open pull requests of the child sessions archived alongside it. Devin also became a better on-call citizen, adding PagerDuty incident automations and 21 one-click OAuth MCP marketplace integrations, plus an optional idempotency key on the Sessions API.

Model churn and a retirement deadline

OpenAI gave developers two configuration deadlines. GPT-5.3-Codex-Spark was deprecated outright, disappearing from the ChatGPT desktop app, Codex CLI, and IDE extension, while GPT-5.5 got a longer runway but still needs attention: it retires from ChatGPT, ChatGPT Work, and Codex on October 14, so anyone signing in with ChatGPT should move workspace defaults, saved model settings, managed configurations, custom agents, and scheduled tasks to gpt-5.6-sol. The API is unaffected. On the upside, GPT-6-Astra landed in the model picker and Amazon Bedrock catalogs, alongside experimental worktree support for isolated session checkouts.

Claude Code had a rougher stretch of regressions and recoveries. A gateway sign-in regression that broke proxy and LLM-gateway setups using API keys or custom auth headers was fixed quickly, as was an HTTP 400 failure on third-party Anthropic-compatible endpoints and a WebFetch hang on servers that hold responses open. Read-only git commands in Bash briefly started asking for permission after long sessions.

Plugins get testable, efforts get capped

Claude Code's feature work centered on making extensions first-class. It added claude plugin eval for running a plugin's eval suite with scored JSON and HTML reports, plus /output-style switching across Remote Control and headless sessions, and support for pointing --plugin-dir at a folder of plugins that can be added or removed while running. Alongside came an maxEffortLevel setting to cap reasoning effort across every provider, including Bedrock, Vertex, and Foundry, and a sandbox containment fix for marketplace paths containing backslashes.

The Codex Python SDK reached parity on effort controls, adding max and ultra reasoning-effort values plus an ExternalMessage type for sync and async run() and turn() calls. Note that HookMetadata now nests its handler under .root, so hook.command becomes hook.root.command.

The mobile and terminal long tail

Codex's client surfaces kept pace: ChatGPT for iOS gained @ mentions, background worktree setup via Live Activity, and steadier reconnects, and the desktop app brought quick chats from the floating Pets controls, Windows Appshots, and Codex Micro keys for reusable prompt text. Grok Build shipped a steady stream of polish across four updates, headlined by headless prompts that time out cleanly instead of hanging and correct grok -c session resumption, a single-row session header with a Dashboard button and tmux pane lag fixes, MCP tools with long server prefixes no longer silently dropped, and a Windows on ARM64 TLS handshake crash fix.