An AI agent is a loop: plan, use a tool, read what came back, then decide whether the answer is good enough. Run a simulated agent step by step, break its tools, take away its checks and see where it must stop.
GoalStep 0 / 16needs 6
STEPS USED0 / 16of the step budget
TOOL ATTEMPTS0 / 3retry limit is 3 calls
EVIDENCE0records backing the answer
REVISIONS0 / 1after a failed review
STATEGoalwhere the agent is now
THE TASK
Tool it needs: calculator
INJECT FAILURES
SAFEGUARDS
Changing any setting, including the step budget next to Run, restarts the run from the goal.
THE AGENT LOOP
Where is the agent now?
YOUR CALL · HUMAN CHECKPOINT
Would you send this answer?
STEP BUDGET0 of 16 used
Tap any box to see what that step is for.
Think
Act
Observe
Check
Recover
Stop
THE ANSWER
Nothing yet
In progress
The agent has not called a tool, so it has nothing to say.
Real agents built on language models follow the same shape: decide on an action, call a tool, read the observation, repeat until there is enough to answer. What makes them dependable is not the loop itself but its limits: a retry cap, a review that can say no, a step budget, and a place for a person to step in.
How the model works
One step at a time. Every move between boxes costs one step. The run is a fixed list of rules (a finite-state machine), so the same settings always produce the same trace.
Tool calls. “Tool errors before it works” makes the first N calls fail. After a failure the agent retries. A third failed call stops the run, so 3 errors can never succeed.
Observation. On success the mock tool returns a fixed result: 8 × 25 = 200 credits, a lamp catalog record, or three checklist fields. “No evidence” keeps the result but attaches no source record.
Review. If evidence is required and none is attached, review fails and sends the agent back to the tool once. A second failure stops the run. With review off, the draft is returned marked unchecked.
Human checkpoint. The run waits after drafting until you approve or reject. Approval is remembered, so a later revision goes straight to review.
Step budget. Before every step the agent checks its count. At the limit it stops, wherever it is.
“Needs N” on the budget and timeline comes from a dry run of the same rules with your settings, approving any checkpoint automatically.
HONEST LIMITS
What this is, and is not.
A scripted teaching simulation. There is no language model here, no real tools and nothing is sent anywhere. The “thoughts” in the trace are fixed messages for each state, not generated reasoning.
Assumptions, controls & sources
The review only checks one rule: is evidence attached? Real review (by a model, a test suite or a person) can still miss wrong answers that look supported.
Real agents choose their own next action, so their paths branch far more than this diagram. Retry caps, budgets and checkpoints are the real-world counterparts of the limits shown here.
The tool error message in the trace is a label for the simulated failure; the model records only that the call failed.
Speed settings change animation pace only. One step is one state change, not a measure of time or cost.
Keyboard: Tab to any box in the loop map and press Enter to read about it. Run, Step, Approve and Reject are ordinary buttons.