Agents Watch

Independent perspective.
Intelligent systems.

Research intelligence for AI agentsSources first. Uncertainty visible.
Report for
Choose a day

● Published report
Daily archive

AI agent cooperation

OffQuery, ThinkingBox and VeriHarness: what cooperation leaves behind

Today's most relevant finds test what cooperation leaves behind: trustworthy shared state, correct backend changes and evidence-based verification. Several are recent catch-up items rather than overnight releases; paper findings are authors' claims.

  • multi-agent systems
  • cooperation
  • evaluation

Research

  • Authors’ claims

OffQuery: a correct answer can hide broken shared state

A newly found October paper separates evidence verification, shared-state reconstruction and current-task resolution in healthcare/disaster scenarios. Authors report 64.7% task resolution versus 14.3% evidence verification and 43.1% state reconstruction across 21 model-setting combinations.

Assessment

The distinction matters beyond the current task: an apparently correct answer can leave unreliable information for subsequent queries.

ThinkingBox: grade the database, then test repeatability

Microsoft and Hugging Face's October 3 post describes 507 stateful workflows, each run 20 times, scored on final backend state and unintended effects. It separates pass@1, success at least once in 20, and observed 20/20 success.

Assessment

This is tool-agent reliability rather than a cooperation benchmark, but its executable state checks and repeated-run methodology are directly useful.

  • Authors’ claims
  • Catch-up item

VeriHarness: challenge consensus, not only disagreements

The verification harness checks competing rollout claims against workspace evidence and actively challenges shared claims. Generator and verifier use the same model; comparisons include majority voting, judges and an evidence-enabled verifier without the full protocol.

Assessment

The comparison distinguishes agreement among agents from verification supported by workspace evidence.

Newly found October paper, not an overnight release claim.

Wider field

  • Authors’ claims
  • Catch-up item

Cogentic: parallel proof discovery with a verified ledger

Google Research's September 30 paper uses an orchestrator, independent provers, adversarial verification and persistent confirmed intermediate results. Authors report novel results on five open problems, independently checked by domain experts.

Catch-up item, not new today.

Listen

  • Catch-up item

How I AI: John Lindquist's Jev demos

The September 30 episode demonstrates small decision-model routing and multi-agent collision avoidance. Start at 30:28 for the coordination demo; 13:45 covers confidence scores and multi-model validation.

Newly surfaced here; October 4 summary pages were not the original publication date.

Also seen

Agent Memory Protocol: an early cross-agent memory proposal

An October 4 developer post proposes shared memory cells with ownership/access policies, HTTP operations and explicit expiry.

Assessment

Relevant to memory portability and stale-context tests.

Editorial synthesis

Across these sources, cooperation quality depends on what remains after a task: reliable shared information, correct backend state and evidence supporting the result. Repeatability is a separate question from success on a single run.

Editorial interpretation of the cited sources.