Drill · lap 2 · 2/10 · AGL-103 · Detect a loop that is stuck, not slow

[akash-coded]

2 of 10 · AGL-103 · Detect a loop that is stuck, not slow

agent-loop · medium · implement · ~12 min · no AWS account

A step cap turns a runaway into an expensive runaway. Detecting a repeat turns it into a cheap, diagnosable one. The trick is fingerprinting what the model asked for — and getting the fingerprint exactly right.

def call_signature(message: dict) -> tuple:
    """A hashable fingerprint of the tool calls in one assistant message.

    - One entry per toolUse block: (name, sorted (key, value) pairs of its input)
    - Order-insensitive across calls: the same set of calls in a different order
      is the same signature
    - A message with no toolUse blocks → ()
    """

def is_oscillating(previous: tuple, current: tuple) -> bool:
    """True when the model asked for the same thing twice in a row.
    An empty signature never counts as oscillation."""

The subtle part

get_booking("WRONG") followed by get_booking("XY7Q2M") is not oscillation — that is the model correcting itself, which is exactly what you asked it to do when your dispatcher named the valid arguments. Fingerprint on name and arguments, or your detector kills the run at the moment it recovers.

/drill AGL-103

```python
def call_signature(message):
    ...
def is_oscillating(previous, current):
    ...
```

What this proves

That you can distinguish the same question asked the same way twice from the same tool used twice — and that you know why the difference matters.

---

Submit in this thread

Copy the block from the brief, edit it, and post it as a comment here. The bot replies within a couple of minutes with every check, the skill you demonstrated, and what to do next.

_{Brief in the repo: labs/drills/agent-loop/AGL-103 · next in the sequence: TOOL-102 · how grading works · scoreboard}