AI agent cooperation Instinct
Semantic cooperative games and paired safety-utility scoring
Today's strongest finds are contribution attribution and paired safety/utility evaluation. Some are catch-up items, clearly labeled, rather than all being overnight announcements. Paper results are authors' claims, not independent replications.
Research
Semantic Cooperative Games: contribution without rerunning every team
Defines a Semantic Shapley Value and SLIC, which attributes contributions from one recorded language-flow trajectory. Authors report 93.3% lower computation cost on one medical benchmark; the semantic extraction assumptions need scrutiny.
Assessment
Directly relevant to separating who contributed information from who caused a failure.
Deny Without Disabling: score useful cooperation and safety together
Authorization-paired evaluation requires a team both to block a prohibited use and complete its authorized counterpart. The paper's controlled tests compare reviewing separate fragments with reviewing the combined artifact.
Assessment
This avoids giving blanket refusal a perfect safety score.
Dated September 30, newly found today.
Workspace MCP v2: shared-workspace infrastructure
The October 2 release describes sessionless transport, multi-tenant authentication, tracing, service permissions and 119 tools across 12 Google services.
Assessment
Relevant as infrastructure for realistic agent-cooperation tasks, not a cooperation benchmark or a new MCP specification.
Listen
80,000 Hours: Rob Wiblin on agent swarms and oversight
A short, strongly opinionated episode on collective agent behavior and the limits of monitoring. Jump to 13:45 for organizational cohesion and 14:32 for knowledge carried between swarms.
Recorded September 25; indexed as published October 2.
Also seen
SWARM governance benchmarks
Its documented recipe compares governance regimes using welfare, toxicity, governance cost and multi-seed stability.
Upcoming A2A/MCP orchestration talk
Google's October 2 open-source roundup lists Daryl Ducharme's "Tag-Team Transmission" talk at MCP Dev Summit Toronto on October 6.
Assessment
Worth watching for interoperability lessons.
Unverified leads
Claimed OpenAI MCP Extensions Toolkit release
A secondary October 3 report claims a new 0.1.0 toolkit for ChatGPT UI extensions.
Editorial synthesis
Contribution attribution, authorized usefulness and harmful outcomes address different questions about agent cooperation. The sources in this digest illustrate why a single aggregate score can obscure those distinctions.
Editorial interpretation of the cited sources.