AI Project Manager
The AI Project Manager (PM) is Groove's safety layer for headless agents. It reviews risky operations before they execute, giving you an audit trail without requiring you to watch every agent.
Why It Exists
When agents run in Auto mode, they operate without constant human supervision. Most actions are safe — reading files, writing code, running tests. But some actions carry risk: deleting files, modifying config, running destructive commands.
The PM provides automated review for these moments. Agents do not need to pause and wait for a human. The PM reviews and responds in seconds.
How It Works
- An agent encounters a risky operation
- The agent sends a review request to the PM endpoint
- The PM builds a context package: the operation details, the project map, the agent registry, and scope rules
- A headless Haiku call evaluates the request
- The PM returns a verdict — approved or rejected — with reasoning
The entire cycle takes 1-3 seconds and costs fractions of a cent.
What the PM Reviews
| Operation | Why It Is Reviewed |
|---|---|
| New file creation | Prevents agents from creating files outside their scope or duplicating existing work |
| File deletion | Catches accidental removal of important files |
| Config changes | Guards package.json, tsconfig, CI configs, and other shared files |
| Destructive commands | Reviews rm, reset, drop, and similar operations before execution |
What Context the PM Has
When reviewing a request, the PM has access to:
- GROOVE_PROJECT_MAP.md — the latest Journalist synthesis of project state
- AGENTS_REGISTRY.md — all active agents, their roles and scopes
- Scope rules — the requesting agent's file ownership boundaries
- Operation details — exactly what the agent wants to do and why
This gives the PM enough context to make informed decisions without needing to read the entire codebase.
Permission Modes
Groove offers two permission modes, set per agent at spawn time:
Auto Mode
The PM reviews risky operations automatically. Safe operations proceed without review. This is the recommended mode for most workflows — you get safety without friction.
Full Send
No reviews. The agent operates with full autonomy. Use this when you trust the task completely or when speed matters more than safety.
Always Ask
There is also an Always Ask mode where every action requires human approval. This is available but intended for sensitive environments, not typical development.
Cost
Each PM review uses approximately 2,000 tokens at Haiku rates. At current pricing, that is a fraction of a cent per review. Even heavy usage — dozens of reviews per hour — adds negligible cost.
PM Review Log
The GUI's Approvals tab shows the complete review history:
- Which agent requested the review
- What operation was proposed
- The PM's verdict and reasoning
- Timestamp and token cost
This gives you a full audit trail of every risky operation across all agents.
Fail-Safe Behavior
If the PM encounters an error — network issue, model failure, timeout — it defaults to auto-approve. The design principle is that the PM should never block an agent from working. A missed review is better than a stalled agent.
TIP
The PM is not a security boundary. It is a quality gate that catches common mistakes. For high-stakes environments, use Always Ask mode with human review.