Describe the outcome. Let it iterate.
A workflow needs you to know the steps. A goal does not. State what success looks like, and AGNT plans the work, runs it, grades the result against your criteria, revises the plan, and goes again — until it passes or it tells you plainly that it could not.
Plan → Execute → Evaluate → Replan → Repeat
This is the difference between an agent that tries once and an agent that gets there. Each iteration is recorded, so you can watch the reasoning improve rather than just receiving a final answer.
Plan
The goal is decomposed into concrete, ordered tasks with dependencies made explicit.
Execute
Tasks run against real tools — files, APIs, databases, shell, other agents.
Evaluate
Output is graded against the success criteria you defined, not against vibes.
Replan
Failed tasks are re-planned with the failure as context, rather than blindly retried.
Repeat
The loop continues until the criteria pass or the iteration budget is spent.
Autonomous does not mean unsupervised
You stay in the loop
Long-running work streams its progress. You can intervene at any point without losing what has already been done.
- Live progress as tasks complete
- Pause and resume without losing state
- Per-task status, output, and errors
- Approval gates on high-consequence actions
Results are graded
A goal produces an evaluation report, not just an artifact — so "did this actually work" has an answer.
- Scored against your stated success criteria
- Written feedback on what fell short
- Automatic, manual, or hybrid evaluation
- Successful runs saved as reusable templates
On honest failure
A goal that cannot meet its criteria reports that it did not. The failure mode we deliberately avoid is the one where a system exhausts its budget and then narrates a success it did not achieve — an evaluation you cannot trust is worse than no evaluation, because it removes the signal that would have told you to intervene.