Official support only. Community ports are mentioned on each entry page.
Counted from each harness's source at the commit researched. Subagent definitions counts agent files only. Some harnesses instead fill in general-purpose subagents from prompt files inside their skills; the entry pages say which.
- Brainstorming
- Has a step that questions you to turn an idea into agreed requirements before planning starts.
- Written plans
- Writes an implementation plan to a file in the repository before coding.
- Test first
- Requires, or strongly directs, writing a failing test before the implementation.
- Debugging process
- Has a defined root-cause process for bugs, as opposed to trying fixes until one works.
- Code review
- Has a review step, usually run by reviewer subagents, before work counts as finished.
- Verification
- Requires running the code or tests and showing the output before claiming the work is done.
- Learning capture
- Writes lessons from finished work to files that later sessions read.
- Branch to PR
- Covers the git end of the work: branches, commits, and opening or finishing a pull request.
- Single entry point
- Has one command or skill that looks at the request and picks the right workflow.
- Automatic activation
- Its skills load on their own when relevant, without the user typing a command.
- Session-start hook
- Uses a hook to inject instructions at the start of every session.
- Subagents
- Hands parts of the work to subagents with their own context.
- Parallel agents
- Runs several subagents at the same time.
- Git worktrees
- Isolates work in separate git worktrees.
- Multiple hosts
- Officially supports more than one host agent. Unofficial ports do not count.
- Per-project setup
- Can be installed or configured in the repository so everyone on the team gets the same setup.
- Team extensions
- Documents how to add your own skills or override its behavior without forking it.
- Needs nothing beyond the host agent and standard developer tools: no extra API keys, accounts, or hosted services.