🔥 Trending on HN

Claude Projects shifts from storage to coordinated work

3 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
coordinator

The main Claude conversation that assigns work and collects results.

thread

A separate work conversation for one part of a project.

beta

A version still being tested with some users.

What changed

Anthropic, the company that makes Claude, announced a redesigned Projects experience on September 17, 2026. The beta runs inside Claude Code. Projects are no longer mainly folders for files and instructions. A user gives Claude a goal, plus a repository or other context. Claude scopes the request, delegates tasks, coordinates parallel threads, reviews the outputs, and assembles a result. The official Anthropic article describes this new workflow.

The story also appeared on Hacker News. The supplied listing shows 78 points and 75 comments. Those numbers measure community attention. They do not prove that the product works well or that every claim is correct.

Why this was needed

Managing several AI sessions used to require a person to divide the work. The person also had to track handoffs and combine the results. Projects now puts a coordinator in the main project conversation. Users can watch progress there. They can also open one thread and guide its details. Anthropic says users can steer work from a phone. It can keep working after they leave the computer.

Why it matters

This changes the unit of work. The user can describe a project goal, not just ask one question at a time. Claude can then decide which parts can run together. Anthropic gives a performance example. One thread profiles app endpoints. Another tests improvements. Another prepares pull requests. A second example connects API, web, and mobile repositories. Each thread migrates callers away from an old endpoint. Claude then reports which changes should merge first.

Parallel work can reduce waiting when tasks are independent. It cannot remove dependencies. If two threads change the same code, their work may collide. The person still needs to set the goal, check assumptions, review changes, and decide what to merge. The coordinator moves this work into one conversation. It does not make human judgment unnecessary.

What is confirmed

Under the design, every thread is a full Claude Code cloud session. It uses its own branch and copy of the repository. Overlapping changes become a normal merge conflict, like a conflict between pull requests. A thread can also split its task into subagents, loops, and workflows.

Projects add shared memory. Threads can add to, and draw from, information about the project. A library collects files and artifacts produced during the work. Users can choose the model and effort level for the coordinator and worker threads. Anthropic warns that several threads can reach usage limits faster.

The beta first opened to selected Pro and Max subscribers using cloud sessions in Claude Code. Access was limited for people without existing web or desktop projects. Anthropic said access would expand, with updated Projects for broader Claude, Team, and Enterprise plans later. Threads run in the cloud today. Local work beside a user's tools and code is described as coming soon.

What remains unknown

The announcement does not provide a general speed test. It gives no broad measure of output quality, merge effort, or total usage for typical projects. It is also unclear which tasks will benefit most after the beta expands. The timing for local execution is not fixed. A coordinator may reduce context switching, but it may also create more work when tasks are poorly divided.

What to watch next

The useful test is not whether Claude can start many threads. It is whether people can understand their progress and review the combined result. Watch the rollout, project usage controls, local access, and conflict handling. Parallel threads look most useful when tasks are independent, their outputs are easy to check, and a person can still make the final decision.

💬 Debating Claude Code Projects’ Parallel AI Workers

HN commenters saw a promising way to coordinate parallel AI work, but also raised concerns about usage limits, software quality, reviewability, and local/cloud friction.

  • An Anthropic team member explained that in Projects, the user talks to a coordinator, which delegates only the necessary work to multiple threads. The threads run in parallel and send results or questions back to the main conversation; a user can still focus on one thread.
  • Several commenters said the idea resembles home-built orchestrators and subagent setups. They described the main difference as a UI layer that handles inter-thread communication, status, scheduled follow-ups, permissions, and tools, reducing context switching.
  • On cost, Anthropic said its internal A/B tests and evaluations found less than 10% token overhead for the coordinator compared with running sessions directly. Token pacing is not solved; asking the system to aim for about 300k tokens was offered as one workaround.
  • A user self-reported that an answer triggered 7 subagents and exhausted their usage allowance in about 10 minutes. This is an individual reported incident, not a measurement showing that the product always behaves this way.
  • A strong quality objection was that even top models can produce incorrect code, tests that are not genuinely trustworthy, and design compromises that make maintenance harder later. Another user argued that if quality remains acceptable, throughput can be worth it—for example, shipping 40 features in a week instead of 4. That is a user judgment, not a controlled product comparison.
  • On oversight, users worried that once 5 workers run at the same time, it becomes hard to inspect each worker’s diff and executed tests. Relying only on the coordinator’s summary could hide individual mistakes.
  • For local work, Anthropic said Projects can talk to a local machine with permission. Users nevertheless self-reported many confirmation and permission handoffs and found the cloud-first design confusing. The debate is therefore also about cost, quality, oversight, and control over where work runs.

mature digest at 96 comments (revision 2). We fetched 90 comments and sampled 90 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Claude Projects can split one goal into many work threads

📰 Full story: Claude Projects shifts from storage to coordinated work

Anthropic is testing a new way to manage several Claude Code jobs together.

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
coordinator

A main helper that sends out jobs and gathers results.

thread

One work conversation for one job.

merge conflict

A problem when separate code changes do not fit together.

💡 The gist

  • Anthropic, the company that makes Claude, redesigned Projects.
  • Claude can split one goal across several work threads.
  • The beta is limited, and parallel work can use limits faster.

Anthropic announced the change on September 17, 2026. Projects used to work mostly like folders. They held files and instructions. People still had to split jobs themselves. They also had to track each session and join the results.

Now Projects has a coordinator. The coordinator is a main Claude conversation. It receives the goal. Then it sends separate jobs to threads. A thread is one work conversation. Several threads can work at the same time.

Imagine a team fixing an app. One thread checks why the app is slow. Another thread tests a possible fix. A third thread prepares a code change. The coordinator brings their reports together. It can also tell the user which changes should be combined first.

This helps when jobs are separate. People do not need to wait for every small job. However, some jobs depend on earlier work. Other jobs may change the same code. Then the results can clash. A person must check the work and choose what to keep.

Each thread runs as a Claude Code session in the cloud. It has its own copy and branch of the code. A branch is a separate work line. When branches change the same place, they can create a merge conflict. That means the changes do not fit together automatically.

Projects also share project memory. They collect files and things Claude makes. The user can guide the main conversation or open one thread. Anthropic says users can guide work from a phone. The work continues in the cloud after they step away.

The update became a topic on Hacker News. The listing showed 78 points and 75 comments. This shows attention. It does not prove that the feature is useful or correct. The official article says the feature is still in beta.

The first users are selected Pro and Max subscribers. Access will expand later. Local work with a user’s own tools is planned, but no date is given. The important test is simple. Can Claude divide work without creating more checking work?

💬 Claude Code Projects: Parallel AI Workers Made Simple

The idea is useful: one coordinator can divide work among several AIs. But commenters worry about spending, mistakes, and checking the results.

  • Anthropic says the Projects coordinator receives a job, splits it among several AI threads when needed, and brings back their answers or questions. You can also keep working on just one job.
  • Several commenters said this is similar to a work boss, or orchestrator, that they had built themselves. Projects mainly makes communication, progress, scheduled checks, and permissions easier to manage in one place.
  • Anthropic said its internal tests found less than 10% extra token use compared with running sessions directly. It also suggested asking for about 300k tokens as a rough target, but there is not yet a finished way to control overspending automatically.
  • A user self-reported that 7 subagents used up the allowance in about 10 minutes. Quality-focused users worry about wrong code and designs that cause trouble later. Another user said that, if quality is good enough, making 40 features instead of 4 can be valuable.
  • With 5 workers running at once, it can be hard to check each diff and test. Local computer access is possible with permission, but users also self-reported many permission checks and confusion about the cloud-first setup.

mature digest at 96 comments (revision 2). We fetched 90 comments and sampled 90 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Claude gives big jobs to little work chats

📰 Full story: Claude Projects shifts from storage to coordinated work

One big job can become several small jobs.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
thread

A small work chat for one job.

cloud

Faraway computers that do the work.

beta

A version that people are still testing.

Anthropic, the company that makes Claude, changed a place called Projects.

Projects is a place for work.

Claude can split one big job into small jobs.

Each small job gets its own work chat.

This work chat is called a thread.

A helper watches the threads.

That helper is called a coordinator.

The coordinator gathers the finished work.

A person can open any thread and help.

The threads use computers far away today.

That is called the cloud.

Hacker News talked about this update a lot.

That shows attention, not truth.

Projects is still a beta.

A beta is a test version.

Some people cannot use it yet.

Separate jobs may finish faster.

Jobs that touch can bump into each other.

A person still needs to check the work.

💬 Claude Code Projects, Explained Simply

One AI boss asks several helper AIs to work at once. People like the convenience, but worry about cost, mistakes, and checking the work.

  • One AI boss gives jobs to other AIs and brings their answers back. You can also stay with just one job.
  • Anthropic says the boss used less than 10% extra AI words in its own tests, and suggests asking for about 300k tokens as a rough target. There is still no perfect way to slow it down.
  • But a user self-reported that 7 helper AIs used up the allowance in about 10 minutes. More helpers might make 40 features instead of 4, but checking 5 helpers for mistakes is hard.
  • Connecting to a personal computer needs permission. Some users said there were too many checks and that the cloud-based setup was confusing.

mature digest at 96 comments (revision 2). We fetched 90 comments and sampled 90 across the thread. These are HN users’ reports, not independently verified facts.

Sources