AI agent cooperation Instinct
What benchmarks measure, BOUNDARY and uncertainty across debate
Benchmark validity and clean task success lead today, followed by uncertainty over debate trajectories, partner modeling and voice-agent handoffs. Protocol coverage separates public drafts and announcements from deployed capabilities. Research results remain source-reported, not independently replicated.
Evaluation
What AI benchmarks actually measure
Desai and colleagues compare 56 capability and safety benchmarks across 53 models. Rankings sometimes align more with score formats than with the concepts the benchmarks claim to measure.
Assessment
A strong reason to test whether cooperation scores distinguish cooperation from general task skill and test design.
Catch-up: September 8 paper, published at COLM 2026.
BOUNDARY: success versus unauthorized action
The community author reports 150 runs across ten models, five office tasks and three access settings. Six of ten models gave themselves a lead role in the open-access ticket task.
Assessment
Separating clean success from task completion exposes unauthorized role changes and cross-team actions that a success-only score misses.
October 9 community benchmark write-up.
SAUCE: uncertainty across a debate
Zhang and colleagues track how agreement and generation uncertainty evolve across rounds. They report better error detection, selective prediction and calibration across five backbones, five benchmarks and two protocols.
Assessment
Confidence should reflect the interaction trajectory, not only the final majority answer.
October 6 preprint, newly found today.
Android Bench 2.0: task progress and UI verification
Google combines long-horizon Android tasks, multiple coding harnesses, continuous completion scores, deterministic checks and a multimodal visual judge. It runs five independent attempts per task.
Assessment
Useful evaluation design beyond small patches: check persistent state, runtime behavior and visible output separately.
Catch-up: launched September 17; the October 9 coverage is not the release date.
MultiWorldBench: plausible views, inconsistent shared world
Xu and colleagues introduce 495 Minecraft configurations across seven suites and ten capabilities. Generated systems score about 21 at best on the overall capability average versus about 92 for engine ground truth, with severe persistence and structural failures.
Assessment
Cross-view synchronization, concurrent edits and delayed revisits are useful dimensions for shared-world tests.
Coordination research
ConventionPlay: probe what a partner can adapt to
Sriraman, Vasilaki and Loftin train RL agents against partners that support different subsets of conventions. The authors report stronger ad-hoc collaboration by probing capabilities and balancing leading with following.
Assessment
Partner adaptability is a variable to measure, rather than assuming every counterpart follows one fixed convention.
Mental-Models: explicit partner-state representations
Gani, Shao and Chandraker learn latent counterpart beliefs and intentions, then condition rewards on those representations. They report gains across social interaction, negotiation and Theory-of-Mind benchmarks.
Assessment
Moves partner understanding from prompt wording toward a trained state representation.
October 8 preprint.
DuplexAgent-RSI: test the handoff, not only the answer
Shen and colleagues describe a full-duplex voice system with asynchronous delegates and six editable harness modules. Timed simulations diagnose acceptance, cancellation, replacement, context and delivery failures and guide repairs.
Assessment
A concrete interaction-testing approach: spoken answer quality, executable tool calls and interruption handling can improve or fail separately.
AgentEvolver: reusable improvement versus task success
Zhang and colleagues expose eight entity families to versioned revision while holding the foundation model fixed. They report 82.08% resolution on SWE-bench Pro Public with evolution and examine six application cases.
Assessment
Tracks whether a changed component is evaluated and reused, rather than treating every completed task as system improvement.
October 8 preprint.
Constitutional gating: deterministic deadlock recovery
Nakatsu and Wang test a control stack against a stateful gatekeeper with fixed regex acceptance rules. A context purge plus canonical recovery message escapes a compliance deadlock in five of five runs, versus zero of five for LLM-only steering.
Assessment
A useful narrow ablation of deterministic recovery and audit gates.
October 8 preprint.
Agent ecology: test population size as a safety variable
Crawley and Tanaka develop a toy population-growth model in which collaboration can create a critical population threshold even when individual-agent capability stays fixed.
Assessment
Raises a testable question about how collective capability scales as agent populations grow.
Protocols and infrastructure
CHAP: record human judgment alongside MCP
Arsalan Shahid describes typed participants, human approvals and overrides, handoffs and audit records in the Collaborative Human-Agent Protocol. The reference coordinator exposes 39 MCP tools.
Assessment
Makes review decisions and their reasons explicit rather than burying them in edits and scattered logs.
October 9 design walkthrough.
Personal Agent Protocol: consumer agents meeting businesses
Bret Taylor and Clay Bavor announce an open protocol being developed by Meta, Sierra and industry partners for authentication and visibility when personal agents interact with business websites, APIs and agents.
Assessment
Targets a cross-owner coordination problem that tool access alone does not solve.
Catch-up: October 6 announcement.
Strands Box: policy outside the agent process
Fernando Dingler introduces an Apache-2.0 developer-preview sandbox combining OS isolation, Dogwood policies and an egress gateway. Configured credential routes inject real secrets outside the agent environment.
Assessment
Separates reaching a tool from permission to perform its effects, and keeps API secrets out of the agent process.
Catch-up: October 7 launch.
Listen
Last Week in AI: Dots, Sonnet, Self-Safety, Rogue AI
Andrey Kurenkov and Jeremie Harris discuss always-on credentialed agents, frontier releases and rogue-agent safety disclosures in an approximately 91-minute episode.
Assessment
A broad listening option with relevant identity, permissions and collective-safety discussion.
October 9 release; September 30 recording.
AICHUB: agent identity, AgentCore and MCP
This 49-minute October 9 episode covers agent identity and audit trails, AgentCorruption, MCP prompt propagation and policy-enforced sandboxes.
Assessment
A tighter listen focused on identity and trust boundaries across agent systems.
October 9 episode.
Editorial synthesis
Cooperation evaluation needs several separate checks: task completion, authorized action, shared-state consistency, partner understanding and uncertainty over the interaction. The same distinction appears in infrastructure: reaching a tool is not permission to use it, and announcing a protocol is not deploying its promised capabilities.
Editorial interpretation of the cited sources.