Anthropic announced a redesigned Claude Code Projects beta on September 17, 2026. This is not simply a new folder for grouping conversations. Projects add a coordinator and execution threads: the user provides a project goal, Claude scopes the work, delegates tasks, reviews outputs, and assembles the result.
The feature is new and not available to every account. Understanding the problem it solves is more useful than switching products solely because the announcement says "multi-agent."
How the redesigned Projects work
When creating a Project, the user selects a goal and repository or other context. A Project can also be configured with a cloud environment, connectors, plugins, instructions, and a model. The coordinator then plans the work while separate threads execute tasks.
Anthropic's examples include:
- Profiling multiple endpoints, testing latency improvements in parallel, and opening pull requests.
- Connecting API, web, and mobile repositories to remove calls to a deprecated endpoint.
- Assigning implementation, testing, and review to separate threads before reporting the required merge order.
The difference from manually opening several Claude Code sessions is coordination. The coordinator manages dependencies and combines the work. It does not take ownership of architecture decisions: an unclear goal can simply produce more changes that need to be redone.
Who has access?
At announcement, the feature was in beta and initially available to selected Claude Pro and Max subscribers who use Claude Code cloud sessions and do not have existing Projects on web or desktop. Anthropic said access would expand gradually, with broader Claude, Team, and Enterprise support later.
If the new interface is missing, the setup is not necessarily broken. Two customers on the same named plan may receive a staged rollout at different times. The account interface and current official rollout notice are the authoritative indicators.
Projects versus subagents
A subagent is a specialized worker created by a main agent for a bounded task. Projects provide a higher-level container and coordinator for multiple threads, a shared goal, and combined output.
- Use a subagent when one session needs a focused investigation or review.
- Use Projects when work spans repositories or stages and needs long-running coordination.
- Do not split a small bug into several threads merely to use a new feature.
Other important Claude Code changes in 2026
Auto mode
Anthropic made auto mode the default for new Pro, Max, and Team sessions starting August 14, 2026. Enterprise, API, and several cloud-platform environments remained opt-in at the time of the announcement. Auto mode uses a classifier to assess tool-call risk so low-risk work can continue with fewer interruptions.
It is not an absence of controls. Existing permission rules still apply, while overly broad allow rules equivalent to arbitrary code execution are set aside in auto mode. Production secrets, deployment credentials, and database write access still require separate controls.
Artifacts
Claude Code can turn code, connector data, and conversation context into an interactive page such as a pull-request walkthrough, system explainer, dashboard, or release checklist. Artifacts are private to their author by default and can be shared with authenticated members of the same organization. Anthropic's documentation states that they cannot be made public to unauthenticated visitors.
Dynamic workflows
Anthropic described dynamic workflows with Claude Opus 4.8 as a research preview that can plan work and run large numbers of parallel subagents in one session. The announced availability covered Enterprise, Team, and Max plans. It should not be confused with the standard subagent functionality available more broadly.
Good uses for Projects
The redesigned Projects fit large jobs with explicit success criteria: cross-service migrations, coordinated front-end and back-end changes, broad test repair, performance investigations, dependency upgrades, and changes that need multiple pull requests merged in order.
They are a poor fit when requirements are still unsettled, the repository cannot build reliably, automated tests are absent, or nobody can review the output. Fixing the specification and test baseline is more valuable than adding agents in those cases.
Prepare the project before delegating it
- Express the goal as a verifiable outcome, not "improve performance."
- State which repositories may be changed and which are read-only.
- Provide test, lint, and build commands that run automatically.
- Keep production credentials out of the cloud environment.
- Require each thread to report changes, test results, and unresolved risks.
- Have a person review diffs, permissions, and data migrations before merging.
Projects automate some of the coordination among sessions. They do not transfer responsibility for software delivery to the model. They are most credible in repositories with clear boundaries, useful automated tests, and an existing review process.
Related comparison: AI coding agents compared in 2026
Official sources
- Anthropic: Projects redesigned
- Anthropic: Claude Code product and rollout information
- Anthropic: Auto mode in Claude Code
- Anthropic: Artifacts in Claude Code
- Anthropic: Claude Opus 4.8 and dynamic workflows