1. Pick the rulebook
2. The task and its tools
Task: "Book a trip." Tap a tool's name to inspect it. Set each tool as working (available and allowed), blocked (it exists, but a rate limit or policy stops the call), or missing (not in the toolbox). Untick "needed" if the task does not use it.
3. The state machine
Each box is a state written (needed, exists, allowed). The paper's trouble spot is D = (1,1,0): needed, exists, but not allowed right now. The box with the thick outline is where the selected tool is.
4. Your move as the agent
| Action | Allowed by the rulebook? |
|---|
Whole task
5. How often could this bite?
The paper's projection: if each gated call has a 10% chance of being blocked and a task makes 5 such calls, the chance that at least one is blocked is .
What this does not show
- It does not test a real AI model. It runs the paper's rule table, so it shows what the rules allow, not what any model does.
- The 10% and 5-call numbers are the paper's assumptions, not measurements of any platform. Exposure is not the same as an actual deadlock rate.
- Real systems may already have retries or escalation that hide this trap. The paper's proposed experiment has not been run.