Stop Trusting Your AI Agent, Start Conducting It

What it does

TAKT is a CLI workflow engine for AI coding agents. Instead of adding more instructions to prompts and hoping the agent complies, you declare phases—plan, implement, review, fix—as a YAML state machine. Each step gets a narrow persona, permission set, and output contract, and the workflow routes findings back through loops until conditions are met.

The interesting bit

The project treats process enforcement as an external control problem, not a prompt engineering problem. By isolating each task in its own worktree and feeding the agent only the context it needs for the current step, TAKT prevents the polluted context and blurred responsibilities that accumulate during long agent sessions.

Key highlights

• Hard phase transitions via YAML rules (COMPLETE, ABORT, and fix loops) that agents cannot silently skip
• Per-step faceted prompting with distinct personas, policies, knowledge bases, and permissions
• Multi-provider orchestration across SDK-based agents (Claude, Codex, OpenCode) and external CLI tools (Cursor, Copilot, Kiro)
• Isolated worktrees with execution logs and reports, plus a built-in GitHub Action for CI/CD
• Community extension for Spec-Driven Development (takt-sdd) with requirements-to-validation pipelines

Caveats

• Provider setup is split between pure Node.js SDKs and external CLIs; some modes even require tmux
• OAuth permissibility varies by provider and is left for users to verify against terms of service
• The direct “Execute now” mode bypasses worktree isolation and writes changes straight into your working tree

Verdict

Teams running repeated, multi-step coding tasks who need reproducible agent behavior without constant human babysitting should try this. Solo developers looking for a one-off autocomplete should probably keep typing.