Claude Code Projects: Parallel Cloud Agents for Small Dev Teams

What Claude Code Projects actually is

Anthropic's documentation describes a project as one ongoing conversation where Claude coordinates a stream of related work. You paste in a bug report, a stack trace or a list of tasks; Claude answers quick questions in place and starts a thread for anything substantial. A thread is a separate cloud session with its own context window that works on its own branch, opens a pull request, and reports back. The coordinator sees what threads report, not every step they take. An Overview pane groups threads by state - Ready for review, Waiting on you, Working, Landing, Idle and Resolved - beside tabs for the Library, the pull requests and any scheduled routines.

• •Every new thread starts with the project's repositories, project instructions and project memory, so a rule you state once reaches all of them.
• •Project instructions are the standing brief sent to each new thread, capped at 16,000 characters.
• •Project memory is stored as files with a MEMORY.md index that every thread reads when it starts.
• •It runs at claude.ai/code, in the desktop app and in the Claude mobile app - not in the terminal CLI, and not via Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry.
• •Code work needs github.com repositories with the Claude GitHub App installed; GitHub Enterprise Server, GitLab and Bitbucket are not supported for threads.

What Projects is not: parallel agents are not free throughput

Running five threads does not give you five engineers. It gives you five branches that all have to land. Threads watch their pull requests with auto-fix on, pushing fixes when CI fails - but Anthropic's own docs note that GitHub emits no webhook when the base branch advances and creates a merge conflict, so auto-fix cannot react to conflicts. Somebody has to rebase. The same is true of judgement: the coordinator reconciles status, not architecture. Fan-out converts a resource you had plenty of, agent time, into one you never had enough of: senior attention on a diff.

• •Merge conflicts are yours. Threads branch from the default branch and do not know what the others are writing.
• •Review load scales with thread count, not with feature count, because each thread arrives as its own pull request.
• •A thread limit you ask for in conversation is a preference, not a cap; the enforced limit is on threads per day.
• •A thread's sandbox pauses between turns and can resume from a fresh clone, so uncommitted work can be lost - tell threads to commit and push on long tasks.
• •During the beta a project belongs to one user, cannot be shared, and has no organisation-level controls.

The numbers Anthropic has published about parallel agent work

Four figures set the ceiling and the price of this way of working. All four are vendor-published, so treat them as the supplier's account rather than an independent benchmark - but they are specific, dated and checkable, which is more than most AI productivity claims manage.

Solo session, local subagents, cloud threads or a new hire?

Anthropic documents five ways to run Claude Code work in parallel - subagents, agent view, agent teams, dynamic workflows and projects - and notes plainly that running several sessions at once multiplies token usage. The honest comparison for a small team includes the option the docs will never mention: hiring a sixth engineer. Read the last column first. The question is not which approach is most powerful, it is which one matches the shape of this week's work.

How do you decompose work so agent threads don't collide?

Collisions are a decomposition failure, not a model failure. Threads cannot see each other, so two threads asked to touch the same route file will both be right and still conflict. The rule is the one that works with human contractors: one thread owns one vertical slice, end to end, and the interface between slices is agreed before any thread starts.

The review gate: pull requests, CI, auto mode and evals

Threads run in auto mode when the thread's model supports it, which means a classifier reviews actions instead of you, blocking anything that escalates beyond your request, targets unrecognised infrastructure or appears driven by hostile content Claude read. That is a real safety layer, but it is not your release process. In auto mode, pushing to a branch of the repository you are working in and opening a matching pull request run without a prompt, so the human gate has to be somewhere you control: branch protection on GitHub, required checks, and explicit ask or deny rules rather than a sentence in a chat.

• •A boundary you state in conversation, such as 'don't deploy until I review', is honoured by the classifier - but it is not stored as a rule and can be lost when context is compacted. Use a deny rule for a hard guarantee.
• •Deny rules block in every permission mode; add permissions.ask rules where you want a prompt rather than a block.
• •In a project with several repositories, the permission rules and hooks in a repository's .claude/settings.json do not apply to threads at all. Plan for that before you add the second repo.
• •Make CI the gate that holds: tests, lint and a build on every thread's pull request, with branch protection requiring them.
• •Claude Code shipped 'claude plugin eval' in the week of 7-11 September 2026, scoring a plugin against test cases and a no-plugin baseline. Apply the same discipline to any skill your threads rely on.

What does running parallel agent threads cost a small team?

There is no separate compute charge for the cloud VM: project usage counts against the same plan limits as your other Claude Code sessions, and consumes them faster. Anthropic says outright that on a Pro plan you should expect to hit your limit sooner on days you run a project. A thread that reaches the limit waits and continues when the limit resets, so work you left running quietly spends your next usage window. Idle threads are not free either - one wakes and uses your plan again when CI fails or a review comment lands on its pull request.

• •A new project runs every thread on Opus at high effort, which draws on your plan fastest. Drop the thread model or effort for routine work.
• •Claude Opus 5.5, released 22 September 2026 and now the default Opus model in Claude Code, is priced at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20.
• •Anthropic reports Opus 5.5 costs 40% less to run than Opus 5 and generates output more than 30% faster - convert to rupees at the day's rate for budgeting.
• •Context hygiene is a cost control: a follow-up sent to a thread idle longer than the cache lifetime, an hour on Pro and Max, re-reads that thread's whole conversation first.
• •Tell Claude to run two or three threads at a time while you learn the pattern, and to answer small questions itself.

Your first week with Claude Code Projects

Treat week one as a controlled trial on work you would have shipped anyway, not as a migration. The goal is to find out how your codebase behaves when four threads touch it at once, while the blast radius is still small enough to revert on a Friday afternoon.

Signs you're running agent threads wrong

None of the common failure modes are about model quality. They are about ownership, gates, and the gap between what you said in a chat and what your repository actually enforces. Each of the four below has a cheap fix, applied before the incident rather than after.

How Tech Arion helps with parallel agent development

We build software this way every day, so the advice above is what we do rather than a summary of a launch post. Client work arrives as tickets in Ticket Agent, our own AI-native delivery platform. A ticket becomes one scoped unit of agent work, the agent writes the fix on its own branch, the build and tests run, and the change reaches staging with a plain-language summary. A person reviews it and approves the production deploy - always. Nothing automated merges itself into a client's live system. If you are deciding whether to adopt Projects, which work to hand to threads, or how to wire the review gate into your GitHub setup, that is an engagement we run regularly. See techarion.com/services/vibe-coding for AI-assisted product development, techarion.com/services/ticket-agent for the ticket-to-staging workflow, and techarion.com/services/ai-consulting for model selection, pilot design and governance.

Frequently asked questions about Claude Code Projects

The questions Indian engineering leads ask us most often about the Claude Code Projects beta, parallel agent threads and what changes in the day-to-day work of a small team.