Autonomous agent
An agent that operates without step-by-step supervision: given an objective, it plans, acts, evaluates its own results and continues until done or blocked. Autonomy is a spectrum — most production systems bound it with approval gates and budgets.
Autonomy is a dial, not a switch
The interesting question is never "is it autonomous" but "autonomous within what". A useful production agent typically has a bounded objective, a fixed iteration budget, a spend cap, a scoped toolkit and a list of actions it may not take without a human. Inside that box it is genuinely self-directing. The box is what makes that acceptable.
The loop underneath
An autonomous agent runs plan, act, observe, evaluate, repeat. The evaluation step is what separates the useful from the alarming: an agent that cannot assess its own output will loop confidently on a broken approach until its budget runs out. Explicit success criteria, checked each iteration, are what turn the loop into progress.
Knowing when to stop
Termination is a design decision with three answers: the goal is met, the budget is exhausted, or the agent is blocked and needs a human. All three should end in a report rather than silence. An autonomous run that finishes without telling you which of the three happened has not really finished.
In AGNT
AGNT goals run autonomous loops with iteration caps, spend limits and human gates exactly where you place them.
Autonomous agent — common questions
What stops an autonomous agent running up a huge bill?
Iteration caps and spend limits enforced by the runtime rather than requested in the prompt. In AGNT these are configuration on the goal, so the loop halts whether or not the model cooperates.
Can an autonomous agent still ask for human input?
Yes, and the good ones do. Approval gates park the run mid-flight with full context attached and resume once a person signs off — autonomy for the routine, escalation for the consequential.
How is a goal different from a workflow?
A workflow is a defined path you designed. A goal is an outcome you specified, where the agent plans the path itself, evaluates results and replans when a step fails.