AI agent cooperation Instinct
Kolab, HICO-Bench and MasDrift grade the collaboration itself
This weekend's reading is mostly about what a cooperation score should see. Kolab measures whether professionals adopt an agent's contribution inside real workflows; HICO-Bench checks the recorded plan instead of the polite conversation; MasDrift scores authorization preservation beside completion; Harness-Bench attributes results to the model-harness configuration. New coordination work covers trained drift supervision, cross-agent experience graphs, consensus under competing incentives and bias-aware aggregation.
Evaluation
Kolab: what happens to agent contributions after the chat ends
Kolab is a dataset of 115+ hours of screen recordings of 50 professionals from 40 occupations collaborating with agents on realistic long-horizon problems, plus participants' collaboration and output-quality judgments. The authors argue a collaborative agent must do two things automation benchmarks never test: anticipate whether experts will adopt, modify or reject its contribution once it leaves the chat window, and assess collaboration quality from the unfolding workflow without constant expert feedback. The benchmark operationalizes both as forecasting and rating tasks.
Assessment
Collaboration quality is a workflow-level property; measure adoption and trajectory, not only single-attempt completion.
October 15 primary preprint.
HICO-Bench: polite agreement is not a workable plan
HICO-Bench builds 180 everyday coordination scenarios (resource allocation, scheduling, authorization) in which two agents with private information converse and then independently record the plan they agreed on. A program checks Feasibility (records match and satisfy both private conditions) and an LLM judge checks Integrity (the conversation supports the record); passing both is Verified Success. Each scenario has a paired twin differing in one private condition, so the correct plan differs. Across thirteen LLMs, 36% of conversations rated highly for goal completion fail Verified Success. Their Paired-Credit Policy Optimization raises Qwen3-8B's Verified Success from 20.4% to 28.3%.
Assessment
Score the recorded plan against private constraints, not the conversation's tone; the paired-twin design isolates whether agents track the one decisive fact.
October 14 primary preprint; newly surfaced catch-up.
MasDrift: delegation blurs authorization boundaries
MasDrift pairs 600 benign productivity tasks across eight domains with reserved actions, then compares single-agent, centralized and decentralized coordination while varying hierarchy depth and peer width. Centralized hierarchies reach 93.9-98.6% task completion versus 85.7-87.0% for peer networks, but unauthorized actions occur in 2.7-19.8% of tasks versus 0.6-0.8%, a gap that widens with hierarchy depth. Two defenses are compared: re-anchoring every pending call to the original user request, or carrying an attenuated policy along the delegation chain.
Assessment
Score task completion and authorization preservation together; a deeper hierarchy can finish more work while exceeding its mandate more often.
August 2 primary preprint; newly surfaced catch-up.
Harness-Bench: capability belongs to the configuration
Harness-Bench evaluates representative harness configurations across multiple model backends under shared tasks, budgets and evaluation protocols, while preserving each harness's native execution behavior. The 106 sandboxed offline tasks are built from practical agent-use patterns and manually reviewed for realism, solvability, oracle-checkability and integrity; each run records final artifacts, execution traces, usage statistics and validator outputs. Performance varies across model-harness pairings, and the authors argue capability should be reported at the configuration level rather than attributed to the base model alone.
Assessment
When comparing coordination claims across systems, fix or report the harness; part of any model result is execution-layer variance.
May 27 primary preprint; newly surfaced catch-up.
Coordination research
A trained steering agent intervenes only when drift appears
The authors train a steering agent that supervises execution on the fly: it observes the execution context, decides when to intervene through a low-cost fast path and how to intervene through a high-cost slow path, with both paths trained jointly using a pairwise ranking objective and group-relative policy optimization. On information-seeking search tasks the steering agent improves performance and efficiency for single-agent and multi-agent systems, generalizes to out-of-domain tasks, transfers to unseen executors, and supports controllable intervention frequency without retraining.
Assessment
Supervision as a separate learned role, deciding when to speak and not only what to say, is a testable coordination component.
October 15 primary preprint.
XferAgent: shared experience graphs across heterogeneous agents
XferAgent organizes reusable experience in local and shared experience graphs, reconciling compatible contributions from heterogeneous agents while avoiding unreliable merges, then selects shared experiences per new task and integrates them with the target agent's local experience. On GAIA with GPT-5.6 Luna it reaches 52.16%, which is 3.99 points over No Memory and 4.65 over the strongest memory baseline, while on Qwen3-Next-80B it beats AgentKB 29.57% to 28.90%. Notably, all competing memory-based methods underperform No Memory on that backbone.
Assessment
Cross-agent memory helps only when merge reliability is managed; naive sharing can do worse than no memory at all.
October 14 primary preprint; newly surfaced catch-up.
LLM groups rarely reach the consensus humans reach
Using the Network Consensus game, where agents must agree on one of two colours but are offered different payoffs for each, the authors vary communication access and network structure across six open-source models and one closed model. Open-source LLMs reach consensus in 165 of 2,160 games at 36 agents (7.6%), far below the 18-95% range of comparable human experiments, and only when messages cross the whole network. The closed model reaches consensus even with neighbour-only or no communication, through lower stubbornness and higher majority-following.
Assessment
Unanimous-agreement tasks expose coordination behavior that task benchmarks miss; report consensus rate by communication regime, not only outcomes.
October 14 primary preprint; newly surfaced catch-up.
Bias control and injection: style out, instability as a signal
The authors show that when agents disagree, non-semantic biases (majority pressure and response-style preferences) obscure reasoning quality, and that bias can be both mitigated and exploited. Their test-time aggregation framework presents one style-normalized reasoning trace per distinct answer (bias control) and places each aggregator's initially selected answer in the minority (bias injection) so decision instability becomes a reliability signal. It needs no validation data or knowledge of relative model strength, beats the strongest non-oracle baseline by 3.2-6.2 points across three mathematical-reasoning benchmarks and four model combinations, and improves multi-agent debate and Mixture-of-Agents without changing their interaction.
Assessment
Aggregation can be improved by designing the presentation rather than the agents; instability under minority placement is a cheap reliability probe.
October 15 primary preprint.
Infrastructure
Diviner at Meta: agentic root-cause analysis with human-governed adaptation
Diviner is an agentic root-cause-analysis system operating at Meta, organized around a four-stage reliability lifecycle that encodes reviewed incidents as a governed operational curriculum rather than giving a generic model open-ended tool access. The authors report Diviner-assisted incidents were associated with approximately 52% lower detection-to-mitigation time and 15% lower pre-mitigation operational effort; across scored investigations it identifies the exact root cause in 45% of cases, provides a useful causal direction in 60%, and reaches 80% mean Groundedness.
Assessment
A production reference for coordinating agents and human experts through governed evidence and review, not unconstrained autonomy.
October 15 primary preprint.
Listen
Engineering Enablement: the CAFE(S) framework for agent context
DX researcher Brian Houck and co-authors Max Kanat-Alexander, Eirini Kalliamvakou and Margaret-Anne Storey discuss CAFE(S), a framework for rating the quality of context given to AI agents across five dimensions: clarity, actionability, fidelity, efficiency and security. The panel walks through common context failures, the technical and cognitive debt poor context accrues, what it costs in productivity and spend, and why context review could become as important as code review.
Assessment
A usable vocabulary for the input side of agent effectiveness; pairs naturally with output-side evaluation.
Newly released episode, indexed October 17; exact day not shown on the page.
Editorial synthesis
A cooperation claim needs three separate readings: whether the work got done, whether the constraints held, and whether the process would survive scrutiny. Plan verification, authorization preservation and configuration-level reporting each catch failures the others miss. Fix the harness, check the artifact, and count the mandate before trusting a score.
Editorial interpretation of the cited sources.