Anthropic is giving selected Claude Pro and Max subscribers a beta feature that breaks engineering work into parallel Claude Code sessions, reviews the results and assembles them. The important catch is wonderfully unglamorous: running several agents at once can burn through a subscription’s allowance much faster.
Anthropic Watch analysis
What happened
The New Stack reports that Anthropic is redesigning Claude Projects around a coordinator that can take an engineering goal, split it into smaller jobs and delegate them to separate Claude Code cloud sessions. Each worker gets its own branch and repository copy, while developers can monitor, inspect or redirect individual threads.
The coordinator can also use Claude Code’s existing subagents, loops and workflows, meaning one assignment may produce several layers of machine activity. Anthropic says users can share repositories, track project-specific usage and choose the model and effort level used by the coordinator and its workers. Shared memory is intended to carry decisions, project details and working preferences across days or weeks.
A practical example supplied by Anthropic involves retiring an old API endpoint across API, web and mobile repositories. Claude could create separate threads to update callers, run tests and open pull requests, then identify which changes need to be merged. Conflicting edits still become ordinary merge conflicts, rather than being waved away by the magic word “agentic”.
The beta is expanding to more Claude Code users on Pro and Max plans, with mobile support and Team and Enterprise access planned later. Local threads that can reach private networks are also described as forthcoming. Existing Projects users remain on the current version until Anthropic upgrades them.
Why it matters
This is a meaningful change in how coding agents are used. Instead of asking one assistant to complete one task in one session, developers can hand over a goal and supervise a small temporary workforce. That could reduce the tedious business of dividing work, maintaining context and stitching results together.
It also changes the cost and review equation. Anthropic says every parallel thread counts as a full Claude Code session, so speed can arrive carrying a rather large token bill. And separate branches do not make the resulting code correct. The New Stack notes that on Anthropic’s own Real-SWE benchmark, even the top-scoring Fable 5.1 failed more than 60 per cent of the time.
Our read
Claude Code Projects looks less like a smarter autocomplete feature and more like an orchestration layer for software work. That is the interesting bit. The model does not merely write code, it decides how to divide the job and evaluates what comes back.
For developers, the sensible approach is to begin with bounded tasks, watch usage closely and treat every pull request as a proposal from an energetic junior team. Parallelism is useful, but it is not a substitute for architecture, testing or somebody who knows where production is buried.
What to watch
- Whether developers see faster delivery after review and merge time are included.
- How quickly parallel sessions consume Pro and Max allowances in ordinary projects.
- When local execution and private-network access arrive, and what controls they include.
- Whether Team and Enterprise customers receive stronger audit, approval and cost-management tools. Sources & Evidence: The New Stack, 17 September 2026, reporting Anthropic’s feature details and statements from an Anthropic representative. The launch is outside the 36-hour living-story window and is materially distinct from the supplied recent stories, which describe related Claude Code developments rather than this parallel Projects beta.
Discussion spark: Should coding agents be allowed to decide how work is split across parallel sessions, or should developers define the task boundaries themselves before the machines start spending tokens?
Sources and evidence
- Anthropic’s new Claude Code feature could drain your plan before lunch – The New Stack (17 September 2026, 17:01 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.