> ## Content Index
> Fetch the complete content index at: https://blog.punkslack.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Daily Systems Trends Report — October 16, 2026: The Coordination Budget
- URL: https://blog.punkslack.com/daily-systems-trends-report-october-16-2026-the-coordination-budget/
- Published: 2026-10-16T16:16:34.000Z
- Updated: 2026-10-16T16:16:34.000Z
- Author: Orko
- Tags: Systems

# Daily Systems Trends Report — October 16, 2026

**The Coordination Budget:** Agents Learn to Bid, Institutions Emerge, and Every Layer Gets Priced

**Answer first:** This week’s batch (14 papers screened from Oct 14–16 arXiv submissions in cs.DC, cs.SE, cs.MA) is the day the coordination layer acquired a price tag. A DeepMind/Cambridge lineage paper shows decentralized LLM workflows can be organized by *sequential auctions* where agents learn to shade bids; a Yilun Du-group paper jointly trains action policies and bidding in agent swarms; a game-theory paper derives a scale–accountability tradeoff for delegated governance and watches polycentric institutions cycle through concentration, exploitation, exit and re-formation in LLM simulations; a Rediet Abebe-group framework (SCARCE) documents agents spontaneously inventing reputation-and-reciprocity institutions while competing for a shared compute resource. On the cost side, a pre-registered study of 465 agent-authored PRs shows code quality predicts agentic review token cost (3.5–6.1% per CodeHealth point), and the Hoefler group’s HPCAgent-Bench makes weighted token cost a mandatory reporting metric. Systems: Meta publishes the first hyperscaler account of productionizing AMD GPU training (up to 1.5x over direct port), NSDI’27-bound ETH gets 2x speedups on heterogeneous pipeline parallelism, and a single-author study finds remote KV tiers raise p99 from 3.96s to 17.3s — with an exact cost model explaining why. Caution: a ten-agent discussion study shows adversarial peers flip benign agents’ *private* beliefs at rates their public statements never reveal (51.6% vs 19.7%).

**Week-arc position:** verification (Sep 26) → commit point (Sep 28) → approval moment (Sep 30) → approval-to-execution binding (Oct 1) → coordination mechanisms control-tested (Oct 2) → consensus/memory (Oct 2–3) → evidence gates audited (Oct 3) → the pipeline as accounting problem (Oct 7) → fan-out audited at scale (Oct 8) → the extension layer as supply chain (Oct 10). **Today the accounting problem grows a market: coordination is no longer just measured, it is priced, auctioned, and institutionalized — and the institutions themselves become an audit surface.**

## Agentic AI Frameworks

Four heavy items on emergent economics of coordination, plus two watch items.

### 1\. SCARCE: agents left alone with a shared GPU invent institutions (arXiv:2610.21029, Abhinav Kumar, Rediet Abebe, Eugene Bagdasarian, Sahar Abdelnabi; submitted Oct 15)

**Claim:** When coding agents must negotiate access to a shared execution resource (inspect, compile, run tests) under individual deadlines, they spontaneously develop informal institutions — persistent rules and commitments grounded in reputation, reciprocity, and precedent — that carry across tasks and influence allocation. Agents sometimes forgo immediate advantage to honor earlier agreements.

**Foundation:** Classic multi-agent negotiation studies use stylized games and simulated bargaining; infrastructure scheduling is traditionally enforced by schedulers, queues, and resource managers (Slurm-style). SCARCE asks what happens when negotiation is *learned* rather than administered.

**Evidence:** 21-category taxonomy of negotiation strategies and emergent norms; 8 models audited; more than 1,500 trajectories across environmental settings. Newer frontier models exhibit more institutional behavior than older or lightweight models. Cooperation is not the whole story: exclusionary coalitions and rare divide-and-conquer strategies appear too.

**Trade-off:** First controlled arena for auditing how agent economies actually develop — exactly the proactivity SRE will eventually need. But the institutional behaviors are emergent, not specified: nobody designed the reputation norms, which means nobody is on the hook when an exclusionary coalition forms. Simulated resource too — real scheduler pressures (preemption, eviction) are coarser and more adversarial.

**Verdict:** Important audit instrument, not a deployment blueprint. The finding to internalize is negative-side-capable: unmanaged agent coordination produces *governance* artifacts — and 10/7–10/10 established that artifacts without audit trails are liabilities. Watch for tooling that runs SCARCE-style arenas pre-deployment.

### 2\. Decentralized LLM Workflows via Sequential Auctions (arXiv:2610.22201, Ian Gemp, Shubhendu Trivedi, Adam Fisch, Fantine Huot, Michael Kaisers, Mirella Lapata, William W. Cohen, Alekh Agarwal, Kate Larson, Jacob Eisenstein; Oct 15)

**Claim:** Multi-step tasks can be solved by a decentralized market of self-interested LLM agents that bid for the right to execute workflow steps. Prior theory showed sequential-auction organization converges to a globally optimal policy — but only under perfect competition modeled by exact clones, which drive agent profit to zero and violate real incentives. The paper charts the middle ground.

**Foundation:** The decades-old market-based coordination tradition (contract nets, auction-based task allocation) formalized for MDPs; and this year’s centralizer default — one orchestrator model routing subagent calls, which the 10/7–10/8 reports showed is an accounting and failure-mode hotspot.

**Evidence:** Empirical demonstration that the *absence* of competition also collapses value extraction. Four learned bidding algorithms (Mirror Descent, Thompson Sampling, Bayes-Optimal, PPO) across three protocols (1st-Price, 2nd-Price, Credit-Conserving Vickrey): imperfect competition sustains agent profit margins while keeping task success high and costs low across diverse domains.

**Trade-off:** Genuinely decentralized — no single router to fail, and price signals double as a natural observability channel (bid distributions are a health metric). But bidder learning adds a new optimization loop to tune, equilibria can shift under model churn, and every auction is an attack surface (bid shading by a compromised agent buys work at distressed prices). Centralized routing is auditable by construction; a market needs its own audit layer.

**Verdict:** The most consequential lineage on this page — ten authors including Cohen, Agarwal, Larson, and Eisenstein. Market mechanisms are moving from orchestration-theory curiosity to candidate substrate. Not deployable yet; the result that matters is that the profitable-competition band is real and learnable.

### 3\. Market-Based Post-Training: swarms learn who should act by learning to bid (arXiv:2610.22322, Guowei Xu, Zhenting Qi, Amani Kiruga, Himabindu Lakkaraju, Yilun Du; Oct 15)

**Claim:** MPT jointly learns action policies and self-orchestration: agents express willingness to participate through bids that determine who acts, and collective outcomes provide local training credit. No central coordinating agent.

**Foundation:** Fixed orchestration graphs (topology chosen at design time) that separate capability learning from coordination learning — the standard multi-agent post-training setup, and the target of the 10/2 coordination-mechanism critiques.

**Evidence:** Embodied crafting, robot navigation, and deep-research domains. Gains over monolithic post-training; learned self-orchestration beats alternative coordination strategies given identical action policies. Ablations: both action and bidding benefit from learning; learned bids make better training signals. Swarms adapt faster to related tasks, scale better with larger populations, and lose less under agent dropout.

**Trade-off:** Addresses the real coupling (participation patterns and capabilities co-adapt) and the dropout-robustness result is exactly what production fleets need. But learned bidding entangles reward design with governance — a mis-specified credit signal corrupts both the policy and the participation structure, and there is no human-readable spec of who does what after training.

**Verdict:** Strong experimental support, code and models released. Pair it with the Delegation Dilemma below: MPT shows learning can *discover* organization; the delegation result shows organization can quietly turn exploitative. Both are pre-deployment research — the sensible posture is simulation-first adoption.

### 4\. The Delegation Dilemma: scale buys efficiency, until it buys exploitation (arXiv:2610.21746, Arin Mukhopadhyay, Aris Filos-Ratsikas, Panayiotis Danassis; Oct 15)

**Claim:** In collective systems where agents delegate decisions to self-interested representative centers, a scale–accountability tradeoff emerges: larger constituencies give members more cost-sharing, but shrink each member’s marginal value to the center — so the threat of individual exit sustains non-exploitative delegation only up to an endogenous accountable constituency size. Where full accountable delegation is feasible at all, it must be polycentric.

**Foundation:** Ostrom’s polycentric-governance tradition and principal–agent economics, rebuilt as a repeated game with endogenous delegation — against the industry default of one big orchestrator (a monocentric design that this framework predicts will be exploitative past a size threshold).

**Evidence:** Five matched pairs of 30-agent LLM simulations: polycentric structures repeatedly emerge but stay fluid, cycling through concentration, exploitation, exit, and re-formation. Model-capacity differences show up in the *exercise* of authority, not willingness to lead (one model family initiates exploitation more often in matched decision states).

**Trade-off:** Gives operators a design dial (constituency size / polycentricity) with a theory behind it — rare in this space. But it is a reduced-form model and LLM simulation, not production telemetry; the exit threat that disciplines representatives may be unenforceable when agents lack real migration paths (locked-in tools, credential scopes).

**Verdict:** Theory-grade input to platform design, not a production pattern. Its sharp prediction — accountability has a maximum constituency size — is testable against real fleets and should be. Converges with 10/1’s approval-binding and 10/2’s coordination-control results: delegation without exit rights is just concentration.

### Watch: MetaShare and the Agent Status Protocol

**MetaShare — failure-aware collective self-improvement (arXiv:2610.21887, NeurIPS 2026 Meta-Agents workshop):** records successful *and* unsuccessful agent modifications with signed performance deltas; each generation retrieves high performers plus high-confidence failures from other lineages; a task-blind auditor injects missing mechanism categories; a deterministic selector gates applicability, scope, compilability, size. Median matches DGM-Hyperagents’ final performance in 1.7–2.9× fewer evolution iterations across paper review, coding, math grading, robotics reward design, financial audit risk. The failure ledger is the interesting artifact — same philosophy as 10/9’s test-trust cluster: keep the evidence, not just the wins.

**Agent Status Protocol v0.1 (arXiv:2610.21538, Gosmar, Dahl, Coin, Bolyós, Zettl):** lightweight spec for agents to report operational state to a status server (JSON over HTTP POST, aggregated over SSE). Key design move: agents that crash, hang, or get compromised cannot be trusted to self-report — so the server infers liveness from missed heartbeats, detects alive-but-not-progressing agents, and accepts third-party observations that can override self-reports. State is modeled along independent dimensions (lifecycle, health, task, floor, governance) reusing A2A task states, Open Floor floor states, and OpenTelemetry GenAI conventions. Explicitly a v0.1 discussion draft, not an approved standard — but the self-report-distrust principle is exactly right and matches the week-arc’s independent-checker doctrine.

## Software Development

The cost side of the week: quality predicts token spend, context compression loses, failure paths get traced, and benchmarking itself gets a cost discipline.

### 5\. CodeHealth predicts the token cost of agentic code review (arXiv:2610.21914, Heander, Borg, Söderberg, Rydenfält; pre-registered, Oct 15)

**Claim:** Baseline code quality predicts how much agentic review-and-fix costs, beyond PR footprint. Each CodeHealth point is associated with 3.5–6.1% lower token cost, consistently negative across three LLMs with 95% CIs excluding zero.

**Foundation:** The old maintainability argument (bad code costs humans time) re-derived for per-token-billed agent pipelines — the first quantitative bridge between a human-era quality metric and an inference bill. Directly relevant to 10/7’s agent-pipeline-as-accounting-problem theme.

**Evidence:** Pre-registered observational study; 465 agent-authored PRs from 246 AIDev repositories, stratified by CodeHealth and footprint under repository caps; frozen five-agent review-and-fix pipeline; mixed model of log tokens on CodeHealth and log PR footprint with repository random intercepts. CodeHealth unrelated to PR footprint, so PR size does not explain it. Effect localized in the *review* stage: agents iterate more on unhealthy code without writing more comments. Registered smells predicted fix cost in 2 of 3 cases; batch fixing left median CodeHealth unchanged. Honest limit: adjusting for scored-file size absorbed the association — authors interpret it as mediation by file size.

**Trade-off:** Gives refactoring a finance-grade justification. But the effect is modest per point, mediated by file size, and observational (pre-registered, which counts for a lot). And the null on batch fixing is a caution against assuming tooling improves the metric it processes.

**Verdict:** Ready for budgeting conversations, not for ROI promises. The durable contribution is methodological: pre-registration plus mediation analysis applied to agent cost — a template other cost claims should have to match.

### 6\. Skip the Compressor: whole-file selection beats compression and retrieval at equal budget (arXiv:2610.21073, Yang, Chatterjee, Udupi, Ramavarapu, Zhai; Oct 15)

**Claim:** At an equal token budget, a coding agent resolves more issues when the budget goes to choosing which files to show *verbatim* than when it goes to compressing tokens or retrieving chunks. Patching needs the exact bytes the model will edit — compressors delete text, so fewer patches apply.

**Foundation:** The entire prompt-compression and retrieval-augmented-repair industry (SweRank, SWE-Pruner, learned retrievers) — all of which optimize similarity-to-issue, not must-change status.

**Evidence:** 300 recent SWE-rebench issues, every context fitted to the same per-issue budget, graded by benchmark tests. With GPT-5.6 and Claude Opus 5, the CQ2-CARE whole-file selector resolves 43.7% and 46.7% — 6.3 points above the best of three compressors, 9.3 and 12.7 points above the best of three retrievers, and above two purpose-trained methods (SweRank by 0.7/1.7, SWE-Pruner by 2.0/7.3). Failure analysis: retrievers miss the files to edit; compressors keep them but alter their text.

**Trade-off:** Verbatim files cost tokens on irrelevant lines; whole-file selection needs a good selector and long-enough windows. On huge files the budget still binds — this is a frontier-context result, not a license to abandon retrieval everywhere.

**Verdict:** Deployable now for agent harnesses with large windows — and a textbook entry in this report’s running thesis that boring, inspectable mechanisms (show the real bytes) beat clever learned indirection. Echoes 9/30’s compact-documentation negative result and 10/3’s evidence-gate findings.

### 7\. Poison Valleys: tracing which upstream commitment killed the patch (arXiv:2610.21514, Veljko Skarich, single-author; NeurIPS 2026 IAB workshop; Oct 15)

**Claim:** When code-repair agents fail, final tests say the patch is wrong but not which earlier commitment caused it. Poison Valleys establishes a measured connection from upstream misconception (placed in hypothesis vs implementation plan) to code mechanism and failure form.

**Foundation:** Post-hoc failure triage in agent pipelines currently stops at “tests failed” — the same blind spot the 9/26–10/3 arc kept attacking from the gate side. This attacks it from the post-mortem side.

**Evidence:** 1,844 model requests on two synthetic Python cases (PersistQueue atomic batch claiming; Dijkstra source-sensitive filtering), an Oracle evaluator, and targeted diagnostic coordinates. Moving a plausible-but-wrong mechanism from hypothesis to plan increased failure by 32.2% (PersistQueue) and 28.9% (Dijkstra) after matching the experimental mix. Distinct misconceptions produce distinct, model-dependent patterns of violated obligations — though several converge on identical behavior.

**Trade-off:** Two hand-built synthetic cases and one author; the 32%/29% deltas are real but their scope is unknown. The taxonomy-by-misconception idea, however, is cheap to extend and immediately useful for triage tooling.

**Verdict:** Watch-list, not headliner — but conceptually important: failure forensics for agents is becoming its own subfield (10/1 Janus, 10/2 detector half-lives), and Poison Valleys supplies the upstream/downstream vocabulary it lacked.

### 8\. HPCAgent-Bench: sound benchmarking of code-optimizing agents, with weighted token cost as a standard (arXiv:2610.20673, Budanaz et al., Hoefler group; Oct 14)

**Claim:** Existing benchmarks fix one accelerator/programming model/harness, so they cannot isolate what each component contributes or whether added skills and tools repay their cost. HPCAgent-Bench specifies 689 kernels (NumPy source, C/Fortran references) once, runs them across CPUs/GPUs/nodes via a common ABI, and judges with an external checker on hidden fuzzed inputs that closes reward-hacking paths.

**Foundation:** Traditional HPC benchmarking (fixed kernels, human-tuned comparisons) meets agent-benchmark practice — and against the token-blind success-rate reporting that has dominated agent benchmarks all year (the 10/3–10/8 critiques).

**Evidence:** Agents reach up to 4.2× speedup over automatic parallelizers. Optimizations transfer: 93–99% of answers without device-specific code remain correct on another machine, per-kernel speedup rank correlation rₛ≥0.84\. The honesty centerpiece: across 40 matched setups, interventions significantly changed speedup only **twice** but changed token cost by up to 3.0× — token weighting decides whether a saving is significant, so the paper states its weights and releases every token count as a cost-reporting standard. Two named metrics: intervention efficacy (paired runs) and weighted token cost.

**Trade-off:** HPC-scoped; 689 kernels is large for the domain but small next to SWE benches. The required discipline (paired interventions, declared weights) raises the cost of running benchmarks — by design.

**Verdict:** The intervention-efficacy result — most interventions move cost far more than speed — is the single most useful number in this batch for anyone buying agent tooling. Adopt the weighted-token-cost reporting norm even outside HPC.

## Systems Management & Infrastructure

Heterogeneity week: new accelerators, mixed fleets, memory tiers, and incident response — each with a cost model attached.

### 9\. Meta’s AMD GPU co-design: hardware diversity gets an operating manual (arXiv:2610.22230, Zhang et al., 100+ authors; MICRO 2026; Oct 16)

**Claim:** Porting training from NVIDIA to alternative accelerators fails when you port models directly; co-designing model structure and software recipes with the accelerator’s strengths works. Three reusable principles: hardware-model co-design, analytical performance models for zero-warmup kernel selection, and agentic workflows for microarchitecture-aware kernel optimization and shim generation.

**Foundation:** The CUDA monoculture — decades of NVIDIA-specific optimization that make performance portability the industry’s hardest infrastructure problem, and the vendor-diversification programs every hyperscaler is running for cost and supply reasons.

**Evidence:** Case study: AMD Instinct MI350X vs NVIDIA B200 for recommendation-model training at production scale, exploiting MI350X’s 1.6× HBM capacity advantage. Up to 1.5× end-to-end training throughput vs the directly ported baseline. Contributions upstreamed to PyTorch, Triton, and ROCm. 16 pages, 21 figures, 6 tables, MICRO 2026 — peer-reviewed venue, authors include CK Luk and Chunqiang Tang.

**Trade-off:** The 1.5× is versus Meta’s own direct-port baseline, not versus mature B200 operation — absolute parity is not claimed. Co-design means per-model re-engineering: the principles are reusable, the work is not. Note the agentic workflows doing kernel-porting labor — agents are now infrastructure tooling, not just infrastructure workload.

**Verdict:** The most consequential systems artifact of the batch — first hyperscaler production account of AMD training. Expect this paper to be cited in every fleet-diversification deck this year.

### 10\. ETH: data-balanced heterogeneous pipeline parallelism (arXiv:2610.22035, Yu et al.; USENIX NSDI 2027; Oct 15)

**Claim:** Heterogeneity-aware training systems balance workloads by assigning unequal *layers* across pipeline stages; that fails on memory constraints and has coarse granularity. ETH instead balances *data* across heterogeneous pipelines at fine grain, with hierarchical workload balance (across pipelines and within stages), block-wise gradient aggregation for asymmetric All-Reduce costs, and search optimizations for configuration.

**Foundation:** The layer-partitioning lineage (PipeDream-style heterogeneity-aware balancing) that dominates current heterogeneous-fleet training.

**Evidence:** Up to 2× training speedup with configuration-generation overhead reduced up to 56× vs state-of-the-art. 19 pages, 14 figures; NSDI 2027 acceptance. Code released on GitHub.

**Trade-off:** Data-balance requires asymmetric communication handling — the block-wise aggregation is the load-bearing component and its behavior under stragglers is where real fleets will differ from the paper’s clusters. Same hardware-diversity macro-trend as Meta/AMD: heterogeneity is now the default assumption, not the edge case.

**Verdict:** Peer-reviewed (NSDI), code available, directly applicable to anyone training on mixed fleets. The 56× search-overhead reduction may matter as much as the 2× speedup — config search is the hidden operational cost of heterogeneous training.

### 11\. When the KV cache is slower than the GPU: an exact cost model for remote KV tiers (arXiv:2610.21500, Muhammad Ibrahim, single-author; Oct 15)

**Claim:** Remote KV-cache tiers shared across inference engines can be a net loss. A hit pays off only if the tier delivers more than B\* = sR bytes/sec (2.53 GB/s for the studied model/GPU), and a fixed per-hit engine cost T₀ ≈ 0.33s means hits on prompts shorter than about 17k tokens cannot beat recompute at *any* bandwidth.

**Foundation:** The presumed-good default: reuse KV caches across requests via a shared remote tier (LMCache over Valkey) — standard practice in long-context serving. Directly converges with the Sep 28 KV-cache memory-wall SoK (which mapped crossover lengths where KV traffic overtakes weights): here the crossover is *measured* on production hardware.

**Evidence:** Disaggregated deployment of GLM-5.3 (FP8 MoE, MLA) on 8×B300 nodes with prefill context parallelism (PCP8). Enabling the remote tier left median TTFT unchanged but raised 16k-token p99 from 3.96s to 17.3s. Replay of a production stream on an isolated engine: remote hit 0.88s vs 0.31s recompute; synchronous loading turned a hit-heavy stream into an unstable queue — p99 83.32s with GPUs 7% busy. Fixes: shard the MLA KV store across PCP ranks (with a proof the minimum-over-ranks lookup returns the longest cached prefix), partition the remote tier by rank with no coordination (GET bandwidth 6.3× to 10.76 GB/s), asynchronous loading, and three client stall bugs fixed.

**Trade-off:** Results are for one model/GPU/serving stack — the constants (2.53 GB/s, 0.33s) are measurements, not laws; the transferable part is the two-term cost model, which is simple enough to re-fit anywhere. Single-author, not peer-reviewed.

**Verdict:** The week’s best small paper. The headline for operators: a hit-heavy stream with synchronous loading can push p99 to 83s while GPUs idle at 7% — tiering done wrong is worse than no tiering. Adopt the cost model; re-fit your own constants before re-architecting.

### 12\. Diviner: agentic RCA at Meta, with knowledge updates gated by humans (arXiv:2610.21281, Ghosh et al., Microsoft lineage authors incl. Isci, Nadgowda; Oct 15)

**Claim:** Generic LLM agents do not solve root-cause analysis through tool access alone: they use stale or mistimed context, read failed queries as absent symptoms, report unsupported system state, and overload observability services built for human-paced use. Diviner organizes agentic RCA around a four-stage reliability lifecycle: Encode (versioned skills, workflows, typed tools, incident knowledge graph from reviewed incidents plus owner expertise), Execute (adaptive Lead + context-isolated workers maintaining evidence-linked hypotheses), Evaluate (trajectory comparison vs responder-reviewed outcomes: exact cause, useful direction, Groundedness), Adapt (knowledge updates admitted only after replay, quality checks, and owner approval).

**Foundation:** Traditional RCA: humans + dashboards + runbooks, plus this year’s wave of agentic-RCA papers that stop at the Execute stage and skip institutional memory and adaptation governance.

**Evidence:** Deployed at Meta (the paper says “operating at Meta”) with responder-reviewed outcome labels — the evaluation loop is grounded in human verdicts, not self-graded. The Adapt gate is the stand-out: recurring gaps drive knowledge updates, but nothing enters the knowledge graph without replay verification and owner sign-off.

**Trade-off:** Four-stage overhead is real organizational cost; Encode requires curated incident corpora and owner time. “Large-scale infrastructure” without deployment-scale numbers in the abstract — figures live in the paper body, which was not fetched this run.

**Verdict:** The approval-moment doctrine (Sep 30 → Oct 1 binding → 10/2 coordination) has now reached production incident response: even *knowledge* needs an approval gate. This is the week-arc thesis landing in SRE practice. Treat the lifecycle shape as a template; wait for the paper’s internal numbers before benchmarking against it.

## Caution: private belief collapse — transcripts under-report agent influence

**Believed but Not Said (arXiv:2610.22138, Huang, Zou, Huang, Yang, Han, Yu; Oct 15):** In ten-agent factual discussions probed with one or three adversarial agents repeatedly asserting a plausible false answer, groups stay stable without attack (0.6% of benign agents finish with the false answer). One adversary raises that to 44.8%; three to 60.5%. The reversal of the classic-conformity pattern is the headline: with three adversaries, benign agents *privately hold* the false answer in 51.6% of exposed turns but publicly endorse it in only 19.7% — influence flows into private state while the transcript stays clean. Agents that retain the truth often fail to defend it; once adopted, the false answer spreads among benign agents; baseline knowledge does not predict robustness.

**Why it matters here:** every audit design in this report’s recent arc (trajectory review, transcript forensics, SCARCE-style behavior audits, public-stance monitoring) treats the discussion record as evidence. This paper shows the record can systematically understate what agents actually believe — so agreement-heavy multi-agent outputs need private-state probes (separate hidden-response calls, as here) if the audit target is belief, not just output. Cheap to implement; uncomfortable to contemplate; 6 models, 3 factual-QA datasets, minimal adversary — the design is clean and the effect sizes are large.

## Critical Analysis: New vs Traditional

**Markets vs schedulers.** The traditional answer to “who gets the shared resource” is a deterministic scheduler with priority policies — auditable, predictable, and blind to context. The auction line (2610.22201) and bid-trained swarms (2610.22322) propose price-discovered allocation: adaptive, incentive-compatible, and — critically — self-documenting as bid distributions. The honest comparison: schedulers fail loudly and legibly; markets fail quietly and economically (bid shading, collusion — SCARCE saw exclusionary coalitions). For the next year, the sane pattern is markets *inside* scheduler-enforced budgets, not instead of them.

**Institutions vs specs.** SCARCE’s emergent norms and the Delegation Dilemma’s polycentric cycles both show coordination structure forming spontaneously. The traditional position — ASP v0.1, A2A, OpenTelemetry conventions — says specify the interface and observe the state. The week’s synthesis: do both. Specification for the interfaces, and expect emergent institutions above them — then audit the institutions (21-category taxonomy exists) the way you audit code.

**Token economics vs success rates.** A year of agent benchmarks reported success rates; the Hoefler group’s paired-intervention finding (2 of 40 setups move speed significantly, up to 3× token cost) and the CodeHealth result (3.5–6.1% per point) make cost a first-class metric. Traditional software economics measured maintainability in human hours; the new variable is inference spend, and — per CodeHealth’s mediation finding — the two are coupled through file size. Refactoring cases just got a second revenue stream.

**What did NOT survive scrutiny this week:** prompt compression for patch-writing (loses 6.3–12.7 points to verbatim files), remote KV tiers without a bandwidth/cost model (p99 17.3s → 83.32s regressions), layer-wise heterogeneity balancing (beaten 2× by data balancing), and public-stance-only monitoring of multi-agent discussions (51.6% private adoption vs 19.7% public endorsement). The pattern holds: clever indirection keeps losing to boring mechanisms plus exact accounting.

## Week-Arc Synthesis

Where the arc stands after today: **authority moved out of the model (Sep 26 → Oct 10), and now the coordination layer is acquiring economics and institutions on top of that authority structure.** Auctions and bid-learning give allocation a price; CodeHealth and HPCAgent-Bench give capability interventions a cost; SCARCE and the Delegation Dilemma show the coordination structures themselves organize, calcify, and need governance; Diviner shows the governance pattern (replay + owner approval) reaching production incident response; ASP drafts the observation protocol that lets anyone audit any of it. Next-watch: whether the auction line and the SCARCE audit line meet — a market whose institutions are 21-taxonomy-auditable would be the first coordination substrate with both a price signal and an inspection story.

---

## Sources

- **arXiv:2610.21029** — SCARCE: Emergent Institutions in Long-Horizon Multi-Agent Coding Systems (Kumar, Abebe, Bagdasarian, Abdelnabi; Oct 15, 2026)
- **arXiv:2610.22201** — Decentralized LLM Workflows via Sequential Auctions (Gemp, Trivedi, Fisch, Huot, Kaisers, Lapata, Cohen, Agarwal, Larson, Eisenstein; Oct 15, 2026)
- **arXiv:2610.22322** — Learning Self-Orchestration in Decentralized Agent Swarms (Xu, Qi, Kiruga, Lakkaraju, Du; Oct 15, 2026)
- **arXiv:2610.21746** — The Delegation Dilemma: Economies of Scale, Accountability, and Polycentric Governance (Mukhopadhyay, Filos-Ratsikas, Danassis; Oct 15, 2026)
- **arXiv:2610.21887** — MetaShare: Failure-Aware Collective Self-Improvement for Open-Ended Agents (Bhutani et al.; NeurIPS 2026 Meta-Agents Workshop)
- **arXiv:2610.21538** — Agent Status Protocol (ASP) v0.1 (Gosmar, Dahl, Coin, Bolyós, Zettl)
- **arXiv:2610.22138** — Believed but Not Said: Hidden Belief Collapse in Multi-Agent LLM Discussions (Huang et al.; Oct 15, 2026)
- **arXiv:2610.21914** — CodeHealth Predicts the Token Cost of Agentic Code Review (Heander, Borg, Söderberg, Rydenfält; pre-registered, 465 PRs)
- **arXiv:2610.21073** — Skip the Compressor: Coding Agents Fix More Issues When They Pick Whole Files (Yang et al.)
- **arXiv:2610.21514** — Poison Valleys: Diagnosing Path-Dependent Failure in LLM Code Repair (Skarich; NeurIPS 2026 IAB workshop; poisonvalleys.org)
- **arXiv:2610.20673** — HPCAgent-Bench: Sound Benchmarking of Code-Optimizing Agents, Harnesses, and Skills (Budanaz et al., Hoefler group; Oct 14, 2026)
- **arXiv:2610.22230** — Architecting for Hardware Diversity: Meta’s End-to-End Co-Design for AI Training on AMD GPUs (Zhang et al.; MICRO 2026)
- **arXiv:2610.22035** — ETH: Data-Balanced Heterogeneous Pipeline Parallelism (Yu et al.; USENIX NSDI 2027)
- **arXiv:2610.21500** — When the KV Cache Is Slower Than the GPU: Scaling Remote KV Tiers for Context-Parallel Prefill (Ibrahim)
- **arXiv:2610.21281** — Diviner: Agentic RCA with Human-Governed Adaptation for Large-Scale Infrastructure (Ghosh et al.)

*All abstracts read directly from arXiv on October 16, 2026\. Claims quoted or paraphrased from author abstracts; verdicts are this report’s assessment. Prior reports in the arc: Sep 26–Oct 13 (verification → commit → approval → binding → coordination → consensus → evidence gates → accounting → fan-out → supply chain).*