Eight Hours of Trying the Same Thing

A developer using Amazon Bedrock AgentCore Runtime with the Nova Pro model reported that a CodeInterpreter tool call entered an infinite retry loop: Nova Pro failed to properly acquire the CodeInterpreter session, then kept retrying the same acquisition indefinitely rather than surfacing the failure. Because no API existed to halt the runtime once it started, the reporter warned the loop could run for up to eight hours unattended, and the only way to stop it was to forcibly revoke the underlying IAM permissions out from under the agent. The problem was isolated to Nova models specifically when prompt and tool caching were both enabled; disabling those caching settings resolved it.

Apraxia, plausible at first glance, since the failure centers on a tool (CodeInterpreter) not being used correctly. Ruled out because apraxia describes a single botched attempt at using a tool; here the problem wasn't the attempt itself but what happened after it failed, an unbounded, unmodified repetition of the exact same failing call.

Concrete Thinking, considered, since the underlying trigger was a caching setting interacting badly with session acquisition. Ruled out because there was no ambiguous instruction being followed too literally; this was a retry mechanism with no exit condition, not a misreading of what was asked.

Perseveration, most consistent: once the session-acquisition call failed, the agent (and the surrounding runtime) repeated the identical failing action without adapting, escalating, or stopping, for as long as eight hours, with the only real off-switch being to cut its credentials.

A retry loop with no ceiling is a small bug with an unbounded cost, the difference between a two-second hiccup and an eight-hour runaway job is often just a missing iteration cap or timeout, not a deeper reasoning failure. Any agent given the ability to retry a failed action should also be given a hard limit on how many times it's allowed to do so, and a way for a human to cut it off that doesn't require touching IAM policy.

Set a hard limit on retries and give humans a way to stop the run without touching IAM policy.

Iteration caps; timeouts; operator stop API for runtimes.