---
title: "Claude Projects Turns Claude Into Your Project Manager. The Manager Still Needs Receipts."
description: "Anthropic’s Claude Projects beta can coordinate cloud coding sessions with shared memory. The clever part is routing work. The risk is trusting worker summaries as proof."
date: "2026-09-18"
tags: ["Claude Code", "Claude Projects", "coding agents", "orchestration", "project management", "agent security", "Codex", "Maestro"]
canonical: "https://news.maestromojo.com/news/claude-projects-ai-project-manager-needs-receipts/"
---

**Maestro’s take.**

Anthropic has finally named the real coding-agent bottleneck.

It is not typing code.

It is deciding who does what, remembering why, and noticing which worker is stuck.

Claude Projects hands that job to Claude.

That could be genuinely useful. But the AI manager sees what its worker threads report back—not every step they take. It can coordinate the work. It cannot prove the work is correct by itself.

## TL;DR

On September 17, 2026, Anthropic announced a redesigned Claude Projects beta for long-running work. One continuing conversation acts as the coordinator. You give it bugs, goals, and new requirements. It starts or reuses cloud threads, gives each thread shared project context, tracks their status, and collects the pull requests and files they produce.

This is more than a folder of chats. It is an AI project manager above several Claude Code sessions.

It is also a public beta for selected Pro and Max users. Threads run in Anthropic’s cloud. Code projects currently depend on GitHub. A project belongs to one user. The redesigned Projects beta is not yet available on Team or Enterprise plans, and the beta has no organization-level controls.

Most importantly: the coordinator sees worker reports, not every action in every worker transcript. Treat its status update as a map. Treat tests, diffs, logs, and human review as evidence.

![An origami-style AI coordinator receives status reports from three coding workers while code diffs, test results, and pull requests travel separately through a human review and merge gate.](https://maestromojo.s3.us-west-2.amazonaws.com/fileman/dd2b61c52a194955a85e8919127dabda/site_media/claude-projects-coordination-verification_e78d6750.png)

*Maestro’s recommended operating model—not Anthropic’s official architecture. Coordination keeps work moving. Independent evidence decides what ships.*

## Why Maestro users care

Claude Projects can remove real scheduling and continuity work. One inbox remembers the larger goal, routes new requests, and tracks workers that outlive a single session.

It also moves the verification boundary.

Instead of watching each worker directly, the operator may first see the coordinator’s summary of that worker. That makes durable evidence more important, not less. The new question is not only “Did the worker finish?” It is “What independent result proves the handoff is true?”

That is the useful advance—and the new risk.

## What Claude Projects actually changes

Without Projects, developers start several coding sessions, repeat the background, decide which session owns each job, and check them one by one.

With Projects, the main conversation does that scheduling.

Anthropic’s documentation describes four pieces:

1. **One project conversation.** Claude receives new work and decides whether to answer directly, open a new thread, or route the request to an existing thread.
2. **Worker threads.** Each thread is a separate Claude Code cloud session with its own context and branch. It can open a pull request and keep watching CI and review comments.
3. **Shared context.** New threads receive the project’s repositories, instructions, memory, tools, and cloud environment.
4. **An Overview pane.** Threads are grouped as working, waiting on you, ready for review, landing, idle, or resolved.

The useful change is not “more agents.” Developers already have parallel chats, subagents, worktrees, and cloud sessions.

The useful change is one persistent place that remembers the larger goal and routes the next request.

That is closer to briefing a project lead than opening another terminal.

## The manager has a blind spot

Anthropic says the project conversation sees what threads report back, not every step they take.

That keeps the coordinator’s context manageable. It also creates a trust boundary.

A worker can say the tests passed. The manager can repeat that status. Neither sentence proves the tests ran against the right code.

Project memory has a similar tradeoff. Claude writes memory files as it works. You can inspect and edit them, but later threads may treat those notes as accepted context. A mistaken decision can become durable faster than a forgotten decision.

This does not make Projects unsafe. It means the manager needs receipts:

- the exact test command and result;
- the branch and commit;
- the diff or pull request;
- unresolved warnings and assumptions;
- a human-controlled merge or deployment gate.

A summary helps you find the work. Evidence helps you trust it.

## One multi-repository footnote is not a footnote

In a one-repository Project, threads can apply that repository’s Claude permission rules and hooks.

In a Project with several repositories, Anthropic’s current documentation says the repositories’ `.claude/settings.json` permission rules, hooks, and environment settings do not apply to the worker thread. The repositories’ `CLAUDE.md` files, skills, agents, commands, and plugins still load.

That distinction matters.

Instructions tell the model what it should do. Repository access, network rules, credentials, and merge permissions constrain what it can do.

If you put several repositories into one Project, rebuild the important controls in the Project’s cloud environment and GitHub permissions. Do not assume every repository’s local guardrails came along for the ride.

## The manager has a payroll

Every worker thread is a full Claude Code session. The coordinator uses tokens too. A thread watching a pull request can wake up when CI fails or a reviewer comments.

New Projects default to Opus for both coordinator and threads. Anthropic says threads start at high effort while the coordinator starts at low effort.

You can choose smaller models and lower effort. You can also ask Claude to run fewer threads. But the documented thread limit you state in conversation is a preference, not a hard cap. The enforced ceiling is 200 new threads per day across Projects.

There is another easy-to-miss behavior. When an ordinary project thread hits a plan limit, it can wait and continue automatically after the limit resets. A thread started by a routine is the exception: its turn stops with a limit error and must be messaged after the reset. Unfinished ordinary work can therefore begin using the next allowance unless you stop the thread or pause the Project.

Parallel work saves calendar time. It does not automatically save tokens or review time.

## Claude Projects, Codex, and Maestro solve different layers

| Tool | Main unit | Who coordinates | Where work runs | Durable state | Main tradeoff |
|---|---|---|---|---|---|
| Claude Projects | A long-running initiative | A persistent Claude coordinator routes work to cloud threads | Anthropic cloud sessions on GitHub branches | Project instructions, model-written memory, thread status, files, and pull requests | Elegant one-inbox workflow; currently one user, Claude-only, cloud-first, and beta |
| Codex / ChatGPT | A chat, worktree, or delegated subagent workflow | A parent chat can create and manage subagents; people also run parallel chats in isolated worktrees | Local, connected remote, or hosted environments depending on the workflow | Chat context, repository instructions, skills, and worktree state | Flexible execution and isolation; the cited features do not describe one persistent initiative coordinator equivalent to Claude Projects |
| Maestro | A shared task, approved plan, and activity trail | People and agents coordinate through the same board and role-based workflow | In the user’s connected AI clients, with optional Maestro execution | Board items, plans, decisions, knowledge, comments, commits, and evidence | Cross-provider team record and governance; it is a coordination layer, not the coding harness itself |

Claude Projects is the most opinionated of the three about the top-level interface: keep talking to one coordinator and let it decide where the work goes.

Codex already supports parallel worktrees and subagents. Its parent chat can create, guide, wait for, and consolidate workers. The official OpenAI documentation warns that parallel agents consume more tokens and that writing-heavy parallel work can create conflicts.

Maestro sits above the coding harness. It keeps the task, plan, approval, and evidence outside any one model conversation. That matters when the work spans different AI tools or people.

The products overlap. They are not interchangeable.

## When should a developer use Claude Projects?

**Use it when:**

- one goal will produce several related tasks over days;
- the work can run in cloud sessions against GitHub repositories;
- one person owns the Project and wants one inbox for steering it;
- separate branches and pull requests are a natural handoff;
- shared memory would eliminate repeated setup.

**Do not use it when:**

- one bug fits in one session;
- the work needs a local database, device, VPN-only service, or tool that the cloud environment cannot reach;
- several teammates must jointly own and audit the workflow;
- you need hard concurrency limits rather than remembered preferences;
- you need cross-model or cross-provider routing.

**Why:** Projects removes scheduling work. It does not remove ownership, permissions, verification, or cost.

## One thing to try

Give Claude Projects a small two-repository migration.

Before it starts, tell the coordinator:

- propose the threads first;
- run at most two at a time;
- keep one owner per repository;
- do not merge;
- require each worker to report its commit, exact tests, failures, and remaining assumptions.

Then measure four things:

1. elapsed time;
2. plan or credit usage;
3. human review minutes;
4. defects or missing work found before merge.

Run the same migration once with two sessions you coordinate yourself.

If Projects saves scheduling time without increasing review and repair, the AI project manager earned the chair.

If you spend the afternoon reopening threads and reconstructing what happened, you hired a meeting generator.

## What would change Maestro’s mind

This beta becomes a major workflow shift if real teams show that its coordinator reduces scheduling and handoff work while keeping rework, usage, and escaped defects flat or lower.

The opposite result would matter too. If shared memory spreads mistakes, soft concurrency limits inflate usage, or humans must inspect every thread as deeply as before, then Projects is a cleaner dashboard—not a better manager.

The simple rule:

Let the AI manager route the work.

Make the work bring receipts.

## Sources considered

- Anthropic, [Claude Code Projects documentation](https://code.claude.com/docs/en/claude-projects) — architecture, eligibility, thread behavior, memory, multi-repository settings, limits, usage, and current product limitations.
- Anthropic, [Projects redesigned: from folder to conversation](https://claude.com/blog/projects-redesigned) — September 17 launch announcement and product framing.
- Anthropic Help Center, [What are projects?](https://support.claude.com/en/articles/9517075-what-are-projects) — staged rollout and distinction from the older Projects experience.
- Official OpenAI documentation, [Worktrees](https://developers.openai.com/es-419/docs/environments/git-worktrees) — parallel Codex chats, isolation, and local/remote handoff behavior.
- Official OpenAI documentation, [Subagents](https://developers.openai.com/es-419/docs/agent-configuration/subagents) — parent-agent orchestration, usage, permissions, and the warning about writing-heavy parallel work.
- Maestro Mojo, [Every project needs a Maestro](https://maestromojo.com/) — Maestro’s public description of shared boards, durable task/plan/evidence records, supported clients, and optional execution.

*Maestro’s opinion and analysis were generated with AI, then independently reviewed before publication.*