Key Takeaways

  • Anthropic redesigned Claude Code projects around a coordinator that delegates work to parallel cloud sessions.
  • Shared memory and a project library preserve decisions, files, and working preferences across longer assignments.
  • The model could reduce session management, but teams still need controls for merge conflicts, usage, security, and human review.

Anthropic is turning Claude Code projects from passive containers into active workspaces that can coordinate multiple assignments. Announced on September 17, 2026, the redesigned experience lets a user define a goal while Claude scopes the work, creates parallel threads, reviews their outputs, and assembles the result.

The change addresses a familiar problem in AI-assisted development. Large jobs rarely fit neatly into one conversation. Developers have had to split work among sessions, transfer context manually, and reconcile the output afterward. Anthropic's new coordinator takes on more of that orchestration, while users monitor the main project chat or enter individual threads when closer supervision is appropriate.

Each thread is a full Claude Code cloud session operating on its own branch and copy of a repository. That design follows familiar version-control concepts rather than hiding them. The Git project describes branches as separate lines of development, and Anthropic is applying that structure to concurrent agent work.

For example, a team could ask Claude to reduce checkout p75 latency across several endpoints. Separate threads might profile services, test optimizations, run tests, and open pull requests. In another scenario, Claude could coordinate the retirement of a deprecated v1 endpoint across web, mobile, and backend repositories, then identify the order in which changes should merge.

When two threads modify the same code, Anthropic notes the overlap is handled as a merge conflict, just as it would be with conventional pull requests. The coordinator organizes work, but it does not eliminate dependencies or contradictory changes. GitHub Docs provides the broader operational context: pull requests still rely on review, checks, discussion, and controlled merging. Engineering teams will likely apply the same safeguards when proposed changes come from coordinated agents.

The larger shift is from prompting a coding assistant to managing a small pool of delegated workers. A user can brief Claude much like a chief of staff, specify the repository or supporting context, and configure the environment, connectors, plugins, instructions, and model. Individual threads can divide their own assignments further through subagents, loops, and workflows.

The bottleneck in agentic development is increasingly coordination rather than code generation alone. Producing several patches quickly has limited value if nobody tracks why they were created, which one depends on another, or whether all of them reflect the latest product decision.

Anthropic addresses this coordination challenge through shared project memory. Threads contribute to and retrieve from a common record, allowing Claude to retain details such as a release moving to Friday, a planned export being dropped, or a requirement to consult someone before changing the billing service. Projects also include a library containing uploaded files and artifacts produced during the work.

While persistent context reduces redundant prompting, enterprises must monitor what enters memory, how it is updated, and who can inspect or correct it. A stale decision preserved by the model can spread across several workstreams, requiring teams to treat project memory as operational context rather than an authoritative system of record.

The redesigned projects launched in beta for select Claude Pro and Max subscribers using cloud sessions in Claude Code who had no existing web or desktop projects. Anthropic noted access would expand to more Claude Code users on those plans over the following week, with broader availability across Claude and Team and Enterprise plans coming later. Existing Pro and Max projects were set to continue operating under the earlier model until the wider upgrade reached chat and Cowork.

Usage is another practical constraint. Several concurrent threads mean several complete Claude Code sessions, so projects can consume plan allowances faster. Anthropic provides project-specific usage information and lets users choose model and effort levels for both the coordinator and worker threads. That creates a cost-management question regarding which tasks justify parallel, higher-effort agents, and which are better handled sequentially.

Because the beta targets cloud sessions in Claude Code, organizations evaluating the tool must focus closely on repository permissions, connector access, credentials, logging, and approval policies. Anthropic's Claude Code documentation offers the product-level reference for teams assessing how the coding environment fits their development processes.

The redesign shifts Claude Code toward longer-running software agents, but its business value will depend on governance as much as speed. Teams that define branch ownership, testing gates, review responsibilities, and memory boundaries are more likely to gain useful parallelism without simply creating more pull requests to untangle.