Research note / Aweb

How AI agents work.

An AI agent is not magic autonomy. It is a loop: context comes in, a plan is made, tools are used, results are observed, and a human or system boundary decides what happens next.

02

The Agent Loop

Continuous cycle of intelligent execution

AWeb
THE AGENTLOOP1. PLANUnderstand objectiveand break down2. ACTSelect toolsand execute3. OBSERVECollect resultsfeedback, signals4. REFINEEvaluate outputand next stepHUMAN IN THE LOOPSet objectiveReview progressApprove actionReceive outcome
Agent loop
under operator boundary
1

PlanUnderstand the objective, context, constraints, and what would count as evidence.

2

ActChoose tools, call APIs, write files, or produce the artifact under explicit rules.

3

ObserveRead the result instead of trusting the tone of the answer.

4

RefineCorrect the next step, tighten the boundary, and keep the trace visible.

Human in the loop: set objective, review progress, approve action, receive outcome.
The agent loop is not magic autonomy. It is a repeatable cycle: plan, act, observe, refine, and keep the human operator responsible for the boundary.

The loop matters more than the label.

Many products call themselves agents because they wrap a model in a chat interface. The useful question is different: can the system carry a mission through multiple steps, use the right tools, observe the result, and leave enough evidence for review?

In Aweb terms, an agent is not a personality. It is a bounded workflow that can move through context, decisions, tools, observations, and corrections while keeping the trace visible enough for an operator to inspect.

PerceiveCollect the prompt, context, files, tool results, constraints, and state that matter.
ReasonTurn context into a plan, choose what to do next, and identify where uncertainty remains.
ActUse tools, call APIs, write files, retrieve data, or produce the artifact under explicit rules.
ObserveRead the result of the action instead of assuming it worked because the model sounded confident.
EvaluateCheck the result against the mission, the boundary, and the evidence needed for trust.
RepeatFeed the result back into the next step until the operator has enough evidence to decide.