It is flagged for a loop, with owner, environment, trend, request count, and cap context visible before you click.
Read-only live demo
Follow an agent loop from signal to verified fix.
Start in the agent queue, open the support-bot incident, and trace the full investigation: what Costile detected, which requests proved it, who owned the fix, and whether the runs that followed confirmed the guardrail held.
No signup, setup, or connection required. The demo uses sample production-style data and is safe to explore.
Back to site
Guided investigation
The path through the demo
Review the repeated calls, max-token hits, request timing, and rule that matched the session behavior.
Costile names the concrete change: add a max_turns guard, cooldown repeated turns, and reset context after tool completion.
The follow-up runs stayed below the threshold, giving the team before/after evidence instead of a loose claim.
What you should notice
9
Repeated calls
A support agent re-read its own outputs without a completion limit, so the same work ran nine times inside five minutes.
3 / 9
Max-token cutoffs
Three of those calls stopped at the token ceiling, and that is the evidence pointing to context accumulation as the root cause.
93.4%
Verified recovery
After a max_turns guard was deployed, Costile checked the runs that followed and confirmed they stayed below the guardrail threshold.
$2,935
Projected monthly impact
What the loop would have carried over a month, which is the consequence of the behaviour rather than the reason to investigate it.