Anthropic upgrades Claude Code Projects to run parallel AI workers
TL;DR Highlight
Claude Code Projects now has a coordinator that splits work into parallel threads and opens PRs.
Who Should Read
Developers who run several Claude Code sessions by hand and are tired of splitting work, handing off, and context switching.
Core Mechanics
- Projects moved from folders to conversations
- Set a goal and a repo, and Claude scopes the request, delegates it, coordinates parallel threads, reviews outputs, and assembles the result.
- You can steer from your phone, and work continues after you step away.
- Each thread is a Claude Code cloud session
- Each thread works on its own branch and copy of the repo.
- Overlap in the same code is resolved as a normal merge conflict.
- Threads can split work further with subagents, loops, and workflows.
- Examples: latency and v1 retirement
- For a checkout p75 latency goal, threads profile endpoints, test optimizations, and open PRs in parallel.
- With API, web, and mobile repos connected, it creates a thread per repo and says which PRs to merge first.
- All threads share one memory
- It can remember a release moving to Friday, why an export was dropped, or who to check with before touching billing.
- The article says this reduces the need for complex prompt engineering.
- You can tune check-in frequency and update detail, and a library collects your files and Claude's artifacts.
- You may hit usage limits faster
- Several threads run at once, and each is a full Claude Code session.
- You can see per-project usage and pick model and effort for the coordinator and worker threads.
- Beta, cloud only for now
- It is available to select Pro and Max users on cloud sessions with no existing web or desktop projects.
- Running on your own machine and behind your network is coming very soon.
Community Reactions
- Boris from the team says he uses it daily
- He said Projects is how he writes pretty much all his code.
- He said running many agents in parallel fragmented attention, which motivated the coordinator.
- He said single-threaded deep focus still matters.
- Someone who built a coordinator hit friction
- One commenter built per-microservice expert agents with a coordinator turning todos into threads.
- They said the coordinator session became chaos and needed signal filtering.
- Their biggest issue was the coordinator asking about an idea from an hour earlier and needing a refresher.
- Commenters question what it adds over DIY setups
- One commenter already runs a head orchestrator with sub-orchestrators and a custom MCP spawning other agents.
- They use GitHub issues, worktrees, and PRs to avoid collisions and asked how Projects improves on that.
- Another built a Caldav and XMPP todo pipeline for an LLM.
- Concerns about lock-in and token burn
- One comment argued that more development in the cloud makes leaving harder.
- Several comments said parallel threads burn tokens and raise API costs.
- A comment noted the default model split
- The default coordinator is Opus low and thread subagents are Opus high.
- It said this is the reverse of setups with an expensive orchestrator.
How to Apply
- Turn one parallel job into a goal statement
- Pick today's task and write it as a goal with steps, like reducing checkout p75 latency.
- If you have access, enter the goal in a Claude Code project and see how threads split.
- If not, try the decomposition with the prompt below.
- Plan multi-repo work as a thread per repo
- For API, web, and mobile changes, list per-repo work units and their dependency order.
- Tell each thread to handle PRs and tests, and ask the coordinator for merge order.
- Set coordinator and worker models separately
- In project settings, choose model and effort for the coordinator and threads separately.
- Check the per-project usage view after the first week.
- Feed decisions into shared memory
- Tell the project about schedule changes, why features were cut, and who to contact.
- Set check-in frequency and update detail up front.
Try It
# Paste into Claude or Claude Code (simplified version of the Projects coordinator approach)
You are the coordinator of this project. Do not implement directly; split the work and direct it.
Goal: {{goal, e.g. reduce checkout p75 latency}}
Target repos/docs: {{repo list or documents}}
Answer in this order.
1. Split the goal into 3-5 threads that can run in parallel. For each, give scope, repo/branch, and done criteria.
2. Flag threads likely to conflict on the same code.
3. If threads depend on each other, propose a merge order.
4. List decisions, reasons, and people to check with as a 'shared memo'.
5. Report as {{check-in frequency, e.g. 3 lines whenever a thread finishes}}.
Next step: put each thread description into a separate Claude Code session on its own branch and run them.Terminology
Related Papers
LLM assisted reverse engineering
Pointing an agent at binaries with skills and MD notes made years-long hacks feasible
Metrics-in-the-Loop: Complexity Feedback Reshapes Agent-Written Code
Injecting per-turn complexity reports into coding agents cut code erosion without changing correctness
Why is Gemini 4 Carbon taking so long?
Carbon took nearly an hour on bug fixes that Astra and Opus finished in minutes, mostly from extra unrequested work.
How the Claude Code team thinks about verification
E2E, property and differential tests plus verification rules let Claude run in parallel.
Running coding agents without handing them my credentials
Credentials stay on the laptop; agents run commands there via Touch ID approval
Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use
Reverse engineering shows Windows 11 logs which coding agents you use and reports it as optional diagnostics
The Silent Cost of Conversation History in LLM Reasoning
Keeping full conversation history dropped accuracy from 0.6 to 0.1 within 20 turns; a verified state summary prevented it.
Show HN: AgentRun – open-source, Self-improving Managed Agents
Agents save notes and skills after each run, cutting code review cost to 18 cents