A new multiagent orchestration mode is now in public beta. A lead agent writes a plan that fans out across many agents in phases, then combines results at the end, designed for large and complex workloads.
To enable dynamic workflows, set your agent's type to multiagent_20261001 and ask Claude to run a workflow. Claude will write the plan and orchestrate up to 1,000 agents per run.
In a benchmark on a 116k-line codebase, a single agent found 14 to 27 bugs across three runs. A dynamic multi-agent workflow consistently found 66 of the 70 planted bugs across all three of its runs.
A new guide shows how to deploy a Claude Managed Agent that reads Slack and GitHub on a schedule, maintains credentials and memory, and posts updates automatically. Setup uses a Claude Code command, and the required API credits are included with Max and Team plans.
Every Pro and Max subscriber on the Claude Code Projects waitlist has been let in, accompanied by a short walkthrough video.
A Project is a persistent conversation where Claude coordinates work by running each task as its own parallel thread. Full documentation is available.
A new version of Claude Projects is rolling out that gives users a single Claude agent to coordinate and surface all threads, addressing thread management complaints.
Cat Wu confirmed several user-reported pain points are fixed in the latest app version: folder caps per project, per-project trust levels for prod/SSH, thread location visibility with actionable errors, and remote view access from other Claude instances (phone, web).
MAX plan subscribers now receive monthly Claude API credits ($100 for the $100 tier, $200 for the $200 tier), intended for building personal AI tools on top of the Claude API.
The GitHub repo for an AI-generated homepage Chrome extension is now public after being accidentally shared while private. The extension replaces the new tab page with a daily-generated digest of sites you read, built using the API credits included with Max plans.
A daily-regenerating AI homepage that replaces Chrome's new-tab page, built using Claude Managed Agents and Sonnet 5.5, running in a cloud sandbox for roughly $0.75 per run.
Auto-compact replaces the full conversation with a short summary when triggered. The large token counts people observe are not from compaction itself but from each message accumulating the whole conversation up to that point. Using `/autocompact 400k` keeps per-message token counts lower by triggering summarization earlier.
To turn off auto-compaction: use /config to flip the Auto-compact toggle (sets autoCompactEnabled: false in user settings), or set DISABLE_AUTO_COMPACT=1 at launch for a single session. Manual /compact still works in both cases.
Cache lifetime is 1 hour for regular subscribers and 5 minutes for subagents and API billing. Both can be overridden via `promptCacheTtl` and `subagentPromptCacheTtl` in settings.json.
The `/autocompact` slash command and the `autoCompactWindow` setting in `~/.claude/settings.json` both work on the desktop app, identical to CLI behavior.
Running auto-compact changes the context prefix, which invalidates the prompt cache, but you only pay cache-write cost for the short summary, not the full session. The summary generation itself is mostly cache reads if the cache is still warm.
A user wants the prompt cache to stay warm while the main agent is blocked waiting on a long subagent task. Cat Wu passed the feedback to the team.
Frontier models will periodically scan opted-in open-source projects for vulnerabilities at no cost, with reports including a proof-of-concept, explanation, and suggested fix.
When the Auto mode classifier blocks something a user has already approved, running /permission will unblock it and keep it unblocked going forward.
