# Run a project with Project Autopilot

Canonical HTML: [https://collty.com/help/ai-agents/use-project-autopilot](https://collty.com/help/ai-agents/use-project-autopilot)
Audience: Project owners and delivery leads
Reading time: 18 min
Updated: 2026-07-19T00:00:00.000Z

Turn project evidence into a validated delivery plan, safe autonomous actions and owner-controlled decisions.

## One governed control loop for the whole project

Project Autopilot reads the project's authorized source evidence and produces one coherent delivery plan. It compares the brief and roadmap with the current team, tasks, dependencies, acceptance criteria, cadence, hours, costs and delivery signals instead of reacting to one isolated message.

It does not replace the project owner. It prepares evidence-based options, applies only validated low-risk work allowed by the saved rules and keeps team, budget, deadline, client commitment and invoice decisions behind explicit approval.

## Project Autopilot and PM Twin have different responsibilities

| Agent | Owns | Typical output |
| --- | --- | --- |
| Project Autopilot | Delivery structure and governed change | A validated roadmap, missing tasks, team recommendation, quality findings and approval-ready decisions |
| Digital PM Twin | Ongoing attention and communication | Targeted status questions, follow-ups, owner summaries and client-safe updates |
| Project owner | Consequential authority | Approval or rejection of team, money, timing, client promise and invoice changes |

> **Good to know:** Use both agents only when both responsibilities are needed. Autopilot should not send routine reminders, and PM Twin should not silently rewrite the delivery plan.

## What the review uses

| Evidence | What Autopilot checks |
| --- | --- |
| Brief and project source | Outcomes, constraints, non-negotiable requirements and missing context |
| Architect roadmap | Phases, concrete tasks, dates, dependencies, roles and acceptance criteria |
| Project Workspace | Actual task state, blockers, overdue work, assignment and completion evidence |
| Teams and profiles | Role coverage, verified capabilities, availability, collaboration evidence and capacity risk |
| Delivery policy | Weekly, two-week, monthly or continuous cadence plus time, cost, billing and invoice capabilities |
| Analytics and Team Pulse | Recent delivery movement, workload pressure and collaboration conditions |
| Commercial records | Budget boundary, tracked hours and relevant cost signals when the project enables them |

## Choose the autonomy level deliberately

| Mode | Autopilot may do | Owner remains responsible for |
| --- | --- | --- |
| Draft | Analyze evidence and prepare a complete plan | Approving every proposed action |
| Managed | Create missing tasks only after deterministic validation against saved quality gates | All consequential project changes |
| Within limits | Operate inside saved capabilities, variance and deadline boundaries | Every category marked Always ask approval |

> **Good to know:** Pause is authoritative and immediate. A paused Autopilot cannot run a scheduled review or apply a previously prepared action.

## Review decisions as alternatives, not yes-or-no prompts

1. **Read the reason:** The decision card explains the observed project condition and why owner attention is required.
2. **Inspect evidence:** Check the concrete tasks, dates, role gaps, hours or commercial boundary that support the recommendation.
3. **Compare options:** Each option states its operational impact. The recommended option is identified, but it is not applied merely because the model prefers it.
4. **Choose or dismiss:** Your decision is stored in the project run history. The next review incorporates the selected direction and avoids presenting the same unresolved choice as new work.
5. **Verify the result:** Open Project Architect, Teams Canvas or Workspace from the review when you need to inspect the changed object in its native tool.

## Quality gates run after AI planning

### Inside Collty: Model reasoning plus deterministic validation

The model proposes a structured plan, but server-side rules decide whether it is complete enough to apply. The runtime checks the saved project-agent state again immediately before execution and stores a fingerprinted run trace for audit and later evaluation.

- Missing acceptance criteria, dependencies or role coverage can block application when those gates are enabled.
- The agent cannot invent a specialist profile or silently attach a team; team recommendations use the shared profile-backed assembly engine.
- A material source or project-state change triggers a new review; unchanged evidence does not continuously consume model requests.
- Provider usage, retries and background runs are attributed to the responsible account.
- Later runs compare blocked, overdue, closed-task and Team Pulse movement so outcomes can improve evaluation without treating every correlation as proof.

## Configure the first run

1. **Connect Project Autopilot:** In Agent Center, select the project, add Project Autopilot and open its control panel.
2. **Select autonomy and rhythm:** Start with Managed and Manual or Workdays until the project evidence and approval boundaries are stable.
3. **Choose the team source:** Use My people, Saved teams, External or Best fit according to who may be considered for this project.
4. **Set approval boundaries:** Keep team, budget, deadline, client commitment and invoice changes selected unless governance for the project explicitly permits otherwise.
5. **Save operating rules:** The visible controls are only a draft until the panel confirms Saved.
6. **Run the review:** Read quality findings first, then resolve decision cards or apply a validated task plan.

## If Autopilot does not apply a plan

- Paused or detached agents do not run, even when an old plan is visible.
- Unsaved operating rules do not control the server runtime.
- A quality-gate failure intentionally blocks task creation; correct the source evidence or change the gate deliberately.
- Team, budget, deadline, client commitment and invoice changes wait for owner approval by default.
- A project without enough source evidence should return a bounded request for missing information rather than fabricated detail.
- Smart Console shows new owner decisions, blocked reviews and safe task plans once; open Agent Center for the full evidence.

## Related guides

- [Understand Agent Center](https://collty.com/help/ai-agents/understand-agent-center)
- [Use Digital Pm Twin](https://collty.com/help/ai-agents/use-digital-pm-twin)
- [Plan With Project Architect](https://collty.com/help/projects/plan-with-project-architect)