Anthropic Redesigns Claude Code Projects: Coordinator Agent, Parallel Threads, Shared Memory

Anthropic launched a redesigned Claude Code Projects experience today, replacing what was previously a static file-and-chat organizer with a two-tier multi-agent system: a coordinator agent that accepts engineering goals in natural language and a set of worker threads that execute work in parallel cloud sessions. The beta is available to select Claude Pro and Max subscribers beginning September 17, 2026, with broader rollout planned over the coming week.

Before this release, a developer running a multi-session Claude Code build had to divide the work themselves, pass results between sessions manually, and stitch final output together. Under redesigned Projects, the developer describes what needs done and Claude handles decomposition, delegation, and synthesis.

A Coordinator That Briefs Like a Chief of Staff

The architecture splits into two functional layers. At the top sits the coordinator: a Claude instance the developer briefs in natural language — as Anthropic puts it, "the way you'd brief a chief of staff." The coordinator scopes the incoming goal, routes each piece to a new or existing worker thread, monitors progress, reviews outputs, and assembles the result.

Each worker thread is a full Claude Code cloud session on its own branch and isolated copy of the repository. Threads cannot overwrite one another's in-progress work. When two threads edit the same lines, the result is a standard merge conflict — no new mechanism; Git handles it. Each worker thread can further decompose its assignment using Claude Code's existing subagent and workflow primitives, so a large migration can fan out across files without the coordinator tracking every subagent directly.

Anthropic's launch examples illustrate the pattern: a developer sets the goal of reducing checkout endpoint p75 latency and Claude opens parallel threads to profile each endpoint, test optimizations, and open pull requests. In a second scenario, connecting API, web, and mobile repositories, Claude creates one thread per repo to retire a deprecated v1 endpoint — migrating callers, running tests, and opening PRs before reporting which need to merge first. The developer does not assign tasks, manage session state, or define thread boundaries.

Shared Memory Replaces Per-Session Context Handoff

The most durable addition in the redesign is the memory layer. In single-session Claude Code work, context resets when a session ends. A developer restarting a project session has to re-explain which service owner must approve billing changes, why a feature was dropped, or when the release date moved.

Redesigned Projects ships persistent shared memory that all threads read from and write to. A thread that learns the billing service has a mandatory review gate records that into shared memory; subsequent threads inherit the constraint. A decision logged Monday is available to a thread running Thursday. Alongside memory, a project library collects files the developer uploads alongside artifacts Claude produces — a persistent store rather than per-session output. Anthropic says Claude also learns the developer's working and communication style, including how often to check in and how detailed to make each update.

Read more: Claude Code's /resume command solved the context-loss problem for desktop sessions

How It Compares to GitHub Copilot and OpenAI Codex

Parallel-thread coordination is not new to the competitive landscape. GitHub's Copilot app reached general availability in June 2026 with parallel sessions in isolated local git worktrees and a My Work dashboard aggregating sessions, issues, and PRs. OpenAI's Codex runs up to eight parallel cloud subagents in isolated sandboxes. Both systems are production-deployed.

Anthropic's structural difference: the coordinator is a natural-language conversation. The developer describes an outcome; Claude decides thread boundaries. Copilot requires developers to configure worktrees; Codex requires explicit task definition. Whether that abstraction advantage offsets the most significant gap matters practically.

That gap is cloud-only execution. Each thread runs as an Anthropic-hosted cloud session with no access to local tools, private network resources, or files behind a corporate firewall. GitHub Copilot's worktrees are local — agents run tests against a developer's own environment without sending code to an external server. Anthropic's announcement explicitly acknowledges this: local execution is "coming very soon" but is absent from the current beta. For enterprise teams whose CI pipelines or test environments sit behind a private network, the cloud-only constraint blocks today's most valuable use cases.

Read more: Claude Code's SendFeedback tool and what it reveals about cloud session architecture

Parallel Threads Accelerate Plan Consumption — and an Active Lawsuit

Each worker thread counts as a full Claude Code session for quota purposes. Three parallel threads deplete the rolling five-hour usage window roughly three times faster. Anthropic's announcement explicitly warns that projects will reach usage limits sooner under the coordinator model.

This arrives while a class-action complaint is pending. A proposed class action filed June 14, 2026, in the U.S. District Court for the Northern District of California by plaintiff Karl Kahn alleges that the Max 5x plan delivers approximately 3.5 times the Pro baseline rather than the advertised five times, and that the Max 20x plan delivers roughly six to eight times rather than twenty times. Anthropic's terms of service reserve the right to adjust caps at its discretion; that clause is central to the litigation. Anthropic has moved to dismiss the complaint; a hearing is set for November 6, 2026.

Parallel Projects threads make the contested billing math more immediate. Anthropic's cost documentation, cited by independent researchers, indicates agent teams running in plan mode consume approximately seven times the tokens of a single normal session. The workflow benefit — automated decomposition, persistent memory, parallel execution — is real. The cost is that the effective weekly budget for a multi-thread project is a fraction of what the same plan budget looks like when used sequentially.

Reliability on Real Production Code Remains a Known Gap

Claude Opus 5 leads the vals.ai SWE-bench Verified leaderboard with a 97.0% score as of its most recent update. SWE-bench draws from public GitHub repositories with relatively well-specified issue descriptions, and that benchmark's limitations — overly detailed prompts, single-language bias, contamination risk — are well-documented in the research literature.

The Real-SWE benchmark, which tests agents against private production codebases, shows a starker picture. According to The New Stack's September 14, 2026 analysis, Fable 5.1 scores 38.8% on Real-SWE — meaning it fails more than 60 percent of tasks. The benchmark, developed by Y Combinator-backed Specific Labs, uses licensed private production codebases rather than public repositories. The gap between clean public benchmarks and messy internal codebases is a recurring finding: agents degrade substantially when dropped into environments where context is implicit, tests are incomplete, and relevant business logic is not in the issue description.

The coordinator architecture partially addresses this by inserting a review step between worker output and final synthesis. But reviewer reliability at the coordinator level is not independently established. Developers deploying Projects against production code should treat the review step as a meaningful quality gate, not a solved problem.

Local Execution and Enterprise Access Are the Milestones to Watch

The current beta excludes developers with existing web or desktop Projects — they remain on the current version until Anthropic upgrades them. A waitlist is open for Pro and Max subscribers who do not yet have access. Team and Enterprise access follows the broader rollout, with no specific date published.

The most consequential upcoming feature is local thread execution. Until threads can reach internal test infrastructure, private databases, and network-constrained environments, the Projects architecture is complete enough to demonstrate the concept but limited enough to block the enterprise deployment scenarios where its productivity claims would matter most. Whether Anthropic ships local execution before or after GitHub Copilot consolidates its lead with local worktrees will determine how significant today's beta looks once the rollout completes.