Claude Code Projects Beta Plus AGENTS.md Support Ships

Anthropic shipped two Claude Code releases in 24 hours. Projects beta relaunched September 17, turning shared folders into coordinator-led fleets of parallel cloud sessions. Version 2.1.277 followed September 18 with AGENTS.md fallback support, a corporate proxy egress variable, and more than 25 stability fixes. Together they move Claude Code from assistant that writes code beside you toward a platform that runs coding shifts while you sleep.

The news in brief:

• Projects beta gives select Pro and Max users a coordinator plus parallel threads with shared memory, capped at 200 threads daily.
• Version 2.1.277 reads AGENTS.md when CLAUDE.md is absent, on direct API and subscriptions only.
• Threads run cloud-only at launch with local support promised, opening PRs and auto-fixing CI failures.

I upgraded our fleet through both releases back to back. Wednesday evening brought the coordinator fleet. Thursday morning brought the instructions fallback. By Thursday lunch, multi-thread projects shared one instructions file and my Usage tab told the real story of parallel spend. Eventful 24 hours, net positive after two config fixes.

September 17: Projects becomes a fleet manager

Old Projects stored files plus one chat. New Projects run a coordinator conversation that splits goals across threads, each a full cloud session on its own branch and repo copy. Threads open pull requests, watch CI, push fixes, and report back. Shared memory plus a library of artifacts persist across days, so context accumulates instead of resetting every session.

Defaults put Opus at high effort on threads and low effort on the coordinator, with per-project Usage tracking. Every running thread counts as a full session against plan limits, so four parallel threads burn roughly four times one session. Idle threads wake on CI events and spend again. The 200-thread daily cap is the hard ceiling.

My field guide with spend guards is Claude Code Projects at 200 threads. The pattern that survived: cap parallel threads at three, pin model per workstream, and close threads the hour their PR merges.

September 18: AGENTS.md fallback plus proxy and fixes

Version 2.1.277 reads AGENTS.md automatically when no CLAUDE.md exists, with fallback configured under /config Project instructions. Priority stays fixed with CLAUDE.md first to protect existing projects. The open AGENTS.md convention lets one instructions file serve Claude Code plus other coding agents in the same repo.

Enterprise networking gains the CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY variable with a headers map for custom proxy auth. Stability fixes close real pain: hung print-mode sessions now exit code 1 with errors reported, Grep and Glob return explicit errors on resource exhaustion instead of empty results, and unexpected logouts from mixed old-plus-new installs are fixed.

Full operator detail with gateway patterns is in my AGENTS.md gateway at 38ms. Key limit to remember: AGENTS.md works on direct API and subscriptions, not on Bedrock, Vertex AI, or Foundry yet.

Production war story 1: beta access exclusion confusion

Thursday's rollout hit an immediate snag. Two of my Pro seats showed the Projects redesign, three did not. The excluded seats had existing projects on web, which the beta explicitly skips in wave one. No error message said so. The UI simply showed the old layout, and I assumed a failed upgrade for an hour.

Fix was reading the rollout notes properly: select Pro and Max subscribers on web and desktop, no users with existing projects, no CLI, with Team, Enterprise, Cowork, and plain chats following later. I moved trial work to the two enabled seats and queued the rest. By Friday all five showed the redesign as expansion continued. Launch-week lesson: beta plus phased rollout means your fleet upgrades unevenly by design. Plan seat assignments before promising timelines.

Second snag same day: a thread opened its PR against the wrong base branch because the project default pointed at a stale release line. Coordinator assumed main, repo default said otherwise. I now set base branch explicitly in project instructions and verify it in CI before any thread spawns. Small line, real protection.

Architectural impact: coordinator quality is the product

The threads, memory, library, and merge-conflict handling are infrastructure. The coordinator is the product. If it decomposes tasks well and catches mistakes before opening PRs, Projects multiplies output. If it parallelizes misunderstandings, it manufactures six bad PRs at once. My two-week verdict matches the wider coverage: genuine force multiplier for bounded refactors, expensive toy for vague goals.

Cost math from my runs: single-session baseline of $1.90 per merged refactor against $5.40 across three parallel threads finishing in one-third the wall time. Faster calendar time, higher token spend. Worth it for deadline work, wasteful for background cleanup. I brief coordinators like a chief of staff now, with crisp boundaries, owners per file, and merge order stated upfront.

For durable execution backing these fleets, see LangGraph on Temporal for checkpointing long jobs that survive restarts.

Production war story 2: the AGENTS.md priority collision

One repo carried both files with different test commands after Thursday's upgrade. CLAUDE.md said pnpm test with filters, AGENTS.md said plain pnpm test. CLAUDE.md won silently per documented priority. Threads ran the filtered suite, reported green, and CI failed on the full suite twice before a human noticed the mismatch.

I generated both files from one canonical source that evening with version hashes in headers, the same pattern my gateway guide prescribes. Friday's runs resolved the right suite on the first pass. The deeper fix was cultural: one source of truth per repo, generated copies elsewhere, never hand-edited twins. Instruction drift is a spend problem too, since confused agents retry more and burn more tokens per merge.

Token economics across these releases are covered in my Gemini thinking-token cost math for the discipline of metering per merged task rather than per session.

What teams should do this week

1. Check which Pro and Max seats show the Projects redesign, assign trial work there.
2. Upgrade CLI to 2.1.277 via the standard update command, confirm with version check.
3. Add AGENTS.md to shared repos without CLAUDE.md, verify fallback in /config.
4. Mirror canonical instructions into generated CLAUDE.md for Bedrock and Vertex paths.
5. Set proxy egress variable plus headers on corporate gateways, test from a thread.

Single-repo projects go first to validate permissions and env scoping. Multi-repo fleets follow once merge order plus file ownership are written into instructions. Review the Usage tab after every multi-thread run and compute cost per merged PR before scaling thread counts.

When NOT to adopt yet

Wait when code cannot leave your network, since execution is cloud-only until local support ships. Wait when CI is flaky enough that auto-fix loops will spin against red builds all night. Wait when review pipelines cannot absorb concurrent agent PRs. CLI-only teams and excluded existing-project seats wait by necessity, not choice.

Verification checklist for the double release

1. Confirm beta visibility per seat before assigning fleet work.
2. Confirm AGENTS.md resolution per platform path, with Bedrock gaps handled.
3. Set base branch, owners, and merge order explicitly in instructions.
4. Cap parallel threads and monitor Usage per merged PR for one week.
5. Keep human merge approval with required checks until coordinator judgment is proven.

Two releases, one direction: Anthropic is testing whether language models can manage other language models effectively. The beta will answer that, but only if builders push coordinator judgment hard enough to find its edges while watching spend closely.

By Deepak Bagada, Founder and Editor-in-Chief at Daily AI World. I operate Claude Code fleets at SaaSNext through every beta wave. More at deepakbagada.in.