Four failure classes that show up in every agent

Retry loops, schema drift, state deadlock, and PII leakage. Map them once, catch them everywhere.

Balagei G Nagarajan5 min read

After enough agents, the failures stop looking like surprises and start looking like a short list. Four classes account for most of what we see, and each one is visible from the design, before any code runs.

1. The Retry Loop

An agent calls a tool, the call is flaky, and there is no timeout and no cap on retries. It spins. You can see this the moment a tool with real-world latency sits in the path without a guard around it.

2. Schema Drift

The agent writes to a record whose shape has moved since the plan was drawn. The write fails, or worse, half-succeeds. Any write target that is not pinned by a contract is a candidate.

3. State Deadlock

Multi-turn memory locks up past some turn threshold. Long sessions stall. If the design holds conversation state without a bound, the deadlock is already there.

4. PII Leakage

Private data reaches the model prompt unredacted. The path from a data source to a model call, with no redaction step in between, is the whole tell.

None of these need a running agent to find. They need the goal, the data, and the tools laid out as paths.
— Pavamana AI Labs

None of these need a running agent to find. They need the goal, the data, and the tools laid out as paths. That is what a Faultmap does.


Share

Join the discussion

Have a take, a war story, or a question? Sign in with GitHub to comment and react. Comments are powered by GitHub Discussions, ad-free and yours to moderate.

Continue Reading

Every article here documents a failure mode we've seen in production. Faultmap finds them before you ship.