AI agent cooperation Instinct
MAMUBench and Braid test whether agents work worse together
Two cooperation benchmarks lead today's list. These are newly found October 1-2 papers. Paper results are authors' claims, not independent replications.
Research
MAMUBench / "Worse Together"
Studies agents representing different users with competing demands on shared budgets, calendars and bookings. Compares one coordinator against per-user teams, with and without communication. The study covers 77 scenarios across four environments; the planned benchmark covers 74 across three. Measures group outcomes, fairness and interference, rather than task success alone.
Braid / InFlowOp
A benchmark designed to make agent interdependence necessary: 19 evaluation arms across eight domains, built from ten public datasets. Compares workflows with single-agent baselines under matched turn budgets.
Assessment
Worth checking whether it isolates cooperation quality or mainly rewards decomposition and extra context.
"After Cooperation Is Learned"
Evaluates whether cooperation survives continued training, rather than whether it appears initially. Uses matched cooperative starting checkpoints and time-to-failure analysis.
Assessment
The study treats the persistence of cooperation during continued training as a distinct evaluation question.
"Can AI Scientists Coordinate at Runtime?"
Separates communication, runtime agent selection, and added contracts/verification under matched budgets. Runtime selection has the highest observed mean for each host; adding more coordination machinery lowers those means.
DuoMind / RoboPoly
Adds a long-horizon, distributed multi-robot coordination benchmark, with ablations for semantic communication and hierarchical orchestration.
Assessment
The benchmark extends the study of agent cooperation to embodied, multi-robot settings.
Listen
Latent Space with MIT's Alex Zhang
Persistent subagents at 52:01; swarms at 1:00:30.
Assessment
Relevant to harness design and communication overhead, not a benchmark release.
Also seen
Global Coherence
Distinguishes missing shared information from globally inconsistent actions; includes budget and state-consistency tests.
Fault-Tolerant Budget Conservation
Formal checks and crash/retry experiments for delegation budgets.
Claude Code mods
Official docs confirm event/tool-call hooks suitable for instrumentation.
Network-AI
Repository describes record/replay, comparison tooling and guarded shared-state commits.
MCP Python SDK OAuth vulnerability
A new Oct 2 review led to a verified Sep 28 advisory. HTTP OAuth clients may send credentials to a malicious MCP server's chosen endpoint. Fixed in 1.30.0 and 2.2.0; this does not affect SDK servers or stdio clients.
Older advisory, newly surfaced here, not an overnight release.
Editorial synthesis
These sources highlight several distinct aspects of cooperation: team performance relative to a single coordinator, user-level fairness and consent, and the separate effects of communication and enforcement. Aggregate task success alone may obscure those differences.
Editorial interpretation of the cited sources.