Recovering when an agent’s tool call times out

Fri, 30 Oct 2026 08:55:55 +0000

Genliang Zhu and colleagues study what a tool-using agent should do after a timeout leaves a remote action’s outcome uncertain. A retry can duplicate an effect that already happened or is still in flight.

Their controller considers evidence, waiting, valid deduplication keys, compensation and human handoff. A shared safety check rejects actions that would violate the modeled contract in any compatible state.

On 37 modeled assignments, the planner and three simpler active policies each certified 28 safe completions. Planning saved four synthetic fee units against capability rules while using more computation.
In a separate test of two local models, planner advice certified 30 of 58 paired cases versus nine with evidence prompting. Both conditions used the same safety check. The study did not compare planner advice against advice from simpler rules.
The held-out cases were assignments within previously developed task families. The authors independently replayed trajectories against the same specification; this was not an external replication.

Caveat: The planner receives a supplied finite model of possible provider behavior. The results do not establish exactly-once execution against arbitrary services, or a production cost saving.

Search agents can miss changes they never think to check

Fri, 30 Oct 2026 08:39:22 +0000

Anirudh Phukan and colleagues test whether search-equipped assistants discover recent changes without being told what to check. They pair direct questions with recommendation requests that depend on the same facts.

Eight models were tested on deaths and ended public-office tenures. In native search setups, indirect accuracy on deaths fell by a median 27.5 percentage points at high reasoning effort compared with direct questions.
In the shared search setup, many failures involved no relevant search. Some failures persisted even when a model saw the relevant evidence.
Questions were selected separately for each model using three no-search responses as a proxy for pre-event behavior. The indirect score measures consistency with the targeted update, not the overall quality of an answer.

Caveat: The measured dataset covers deaths and ended tenures. The paper’s museum-closure example illustrates the idea but is not a measured category. The authors propose explicit verification as a possible remedy, but did not test it as an intervention.

Agents on Intelligent Artifact

Recovering when an agent’s tool call times out

Search agents can miss changes they never think to check