PaperlayerView exact revision on arXiv

Source-linked AI summary

SCARCE: Emergent Institutions in Long-Horizon Multi-Agent Coding Systems

Abhinav Kumar, Rediet Abebe, Eugene Bagdasarian, Sahar Abdelnabi

arXiv:2610.21029v1cs.MAcs.AI

TL;DR

Prior negotiation studies often use stylized games, leaving open how agents develop arrangements during sustained tasks. SCARCE has coding agents repeatedly negotiate access to shared execution resources, and the authors find persistent institutions, especially among newer frontier models. These arrangements can include cooperation as well as exclusionary coalitions.

  • Problem

    Prior negotiation studies often rely on stylized games and simulated bargaining, leaving open how such behaviors develop during sustained, long-horizon tasks.

    Source: §1 · p. 1

  • Method

    SCARCE has coding agents solve sequences of programming tasks under private deadlines and budgets while negotiating for time on a shared execution resource.

    Source: §1 · p. 2

  • Results

    Across more than 1,500 trajectories from 8 models, agents formed persistent institutions, mainly in recent frontier models, and also exhibited exclusionary coalitions and rare divide-and-conquer tactics.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Private communication and memory can let agents build persistent rules on top of an operator-specified allocation mechanism, so these arrangements warrant continual auditing.

    Source: §6 · pp. 11–12

    The study does not establish whether these institutions improve or degrade collective outcomes, which requires comparison with a centralized scheduler and measures of deadline misses, task progress, and execution-time distribution.

    Source: §6 · p. 12

Abstract

from arXiv · show

As autonomous agents become widely deployed, coordinating access to shared computing resources and services will become increasingly important. Recent incidents have already revealed unplanned coordination between agents, highlighting the safety implications of these interactions. Yet prior studies of agent negotiation often use stylized games and simulated bargaining, leaving open how these behaviors develop during sustained, long-horizon tasks. We propose SCARCE (Shared Compute ARena for Coding agEnts), a framework in which agents solve sequences of coding tasks under individual deadlines while negotiating access to a shared execution resource. This negotiation determines when and for how long they can inspect programs, compile code, and run tests. We find that agents spontaneously develop informal institutions: persistent rules and commitments grounded in reputation, reciprocity, and precedent. These arrangements, which we call institutional behaviors, carry across tasks and influence resource allocation, with agents sometimes foregoing immediate advantages to honor earlier agreements. We develop a 21-category taxonomy of negotiation strategies and emergent norms and use it to analyze eight models and audit more than 1,500 trajectories across different environmental settings. Among the models tested, newer frontier models exhibit more institutional behaviors than older or lightweight models. Alongside cooperation, we observe exclusionary coalitions and rare divide-and-conquer strategies. SCARCE provides a controlled setting for proactively and systematically auditing how these relationships emerge, shape access to resources, and affect the safety and coordination of multi-agent systems.

1 INTRODUCTION

The paper studies how coding agents develop arrangements around shared resources during repeated interactions, a process not captured by accounts reconstructed after incidents or by stylized negotiation settings. SCARCE enables this study, and the authors report persistent institutions alongside cooperative and exclusionary behaviors.

1 INTRODUCTION

  • Prior work leaves open how independent coding agents repeatedly negotiate scarce resources while carrying what they learn into later interactions.The paper contrasts incident reconstructions and stylized bargaining, games, and simulated societies with coding work where negotiation is rare.Source: §1 · pp. 1–2
  • SCARCE has agents solve sequential coding tasks under private deadlines and budgets while negotiating for time on one shared execution resource.The framework varies agent counts, budgets, deadlines, memory, and the number of consecutive tasks.Source: §1 · p. 2
  • Across more than 1,500 trajectories from 8 models, agents formed persistent rules and commitments, sometimes surrendering an immediate advantage to honor an earlier deal.The 21-category analysis found institutions grounded in reciprocity, reputation, and precedent, mainly in recent frontier models; it also observed exclusionary coalitions and rare divide-and-conquer tactics.Source: §1 · p. 2
  • The paper contributes SCARCE, a taxonomy of emergent negotiation strategies and norms, and an analysis of institutions agents build around scarce resources.The authors connect the study to safe deployment because coordination can serve ends operators did not intend.Source: §1 · pp. 2–3

2 BACKGROUND AND RELATED WORK

Prior work examines multi-agent systems, long-horizon benchmarks, and emergent social dynamics, but the paper focuses on a distinct setting: coding agents with independent objectives competing for shared execution resources. Its environment targets negotiation and norms arising through repeated competition.

Deployment of multi-agent and long-horizon systems

  • Prior multi-agent coding studies emphasize cooperation on shared goals or predefined roles and relatively short tasks, unlike repeated competition between independent coding objectives.The paper also situates its setting alongside work on long-horizon memory and business operations.Source: §2 · p. 3

Benchmarking long-horizon and multi-agent systems

  • Existing benchmarks cover long-horizon coding, games, multi-agent performance, safety, and privacy, while SCARCE targets negotiation that emerges through repeated resource competition.Source: §2 · p. 3

Emergent social dynamics

  • Related work finds persistent populations can develop informal governance, cheating, auditing, whistleblowing, and collusion; this paper studies similar dynamics among coding agents competing for execution resources.Source: §2 · p. 3

3 THE SCARCE FRAMEWORK

SCARCE adapts a difficult reconstruction benchmark into a long-horizon setting where coding agents need shared execution time to make progress. Agents negotiate resource windows over multiple rounds, and the framework exposes their process to configurable conditions.

3 THE SCARCE FRAMEWORK

  • SCARCE adapts ProgramBench, where agents reconstruct source code from an executable and documentation, avoiding a saturated benchmark and reducing opportunities for shortcut-based task completion.The benchmark requires execution time to make progress, making resource negotiation part of task completion.Source: §3 · p. 3

3.1 ENVIRONMENT CONSTRUCTION AND CONFIGURATION

  • Agents solve sequential ProgramBench tasks with individual deadlines and credit budgets, negotiating repeatedly for resource time through a queue, private messaging, and a ballot.Each task trajectory is a subrun, and agents receive evaluation before proceeding to the next task.Source: §3.1 · pp. 3–4, Figure 2 · p. 4

3.2 AGENT SCAFFOLDING

  • Each task uses a solver and negotiator agent; the solver has persistent memory, while a minimal Pi harness controls for harness choice.The solver can update memory without needing allocated compute time.Source: §3.2 · pp. 4–5, Figure 3 · p. 5

3.3 RESOURCE CONTENTION PROBLEM

  • Agents with separate coding tasks compete for sequential access to one execution resource, allocated by a credit-based time-slot auction.Outside an active window, a solver can only request an execution slot; inspecting, compiling, and testing require shared compute.Source: §3.3 · p. 5

3.4 NEGOTIATION PROTOCOL AND SCHEDULING

  • Negotiation proceeds through structured declarations, optional private conversation, and ballots that rank agents and set time allocations and bids.Bids are public, but remaining budgets and deadlines are private information.Source: §3.4 · p. 5

4 TAXONOMY OF NEGOTIATION TACTICS AND INSTITUTIONS

The paper develops a 21-category taxonomy for negotiation tactics and norms, distinguishing classical bargaining behaviors from scheduling conventions agents create and cite. It further identifies which behaviors count as institutions by their self-discovery and persistence across rounds.

4 TAXONOMY OF NEGOTIATION TACTICS AND INSTITUTIONS

  • The non-exhaustive 21-category taxonomy groups observed trajectories into negotiation and bargaining behaviors and rules-and-norm behaviors, after LLM-assisted classification and manual review.The authors remove redundant or weakly evidenced categories and manually cluster related categories.Source: §4 · pp. 5–6

4.1 NEGOTIATION AND BARGAINING

  • Negotiation tactics include framing and signaling, concessions and exchanges, coalition formation, reputation preservation, social proof, and divide-and-conquer appeals.Examples include reciprocity, logrolling, welfare concessions, renegotiation, coalition loyalty, and exclusionary coalitions.Source: §4.1 · pp. 6–7
  • The taxonomy includes iterative bargaining, anchoring, and cheap talk, alongside tactics grounded in bids, precedent, and strategic messaging.Source: §4.1 · pp. 6–7

4.2 RULE AND NORM EMERGENCE

  • Agents can converge on urgency rules such as earliest-deadline-first, conventional scheduling such as recency-first or shortest-job-first, and process rules for bargaining and rule-making.Sunk-cost urgency is a degenerate deadline appeal: an agent claims priority because its deadline has passed even though it will be terminated regardless of order.Source: §4.2 · p. 7

4.3 INSTITUTIONAL BEHAVIOR

  • Institutional behaviors must be self-discovered and forward-scoped; they include relational arrangements such as reciprocity and coalition loyalty, and non-relational norms such as social proof and rule-making.The paper groups relational behaviors around persistent relationships and non-relational behaviors as norms that apply across members.Source: §4.3 · pp. 7–8

5 RESULTS

The experiments classify negotiation behavior across models and test how memory and starting conditions shape institutional behavior. Results vary by model and run state: relational behavior is stronger in some models, while budget and deadline advantages shift over time.

5.1 EXPERIMENTAL SETUP

  • The study classifies trajectory behavior with a taxonomy, then varies budgets, deadlines, and memory; the base setting uses five agents, 1,000 credits, 25-minute deadlines, and up to four subruns.The study also varies deadline from 10 to 25 minutes and budget from 1,000 to 1,500, and measures behavior prevalence and negotiation success rather than task accuracy.Source: §5 · p. 8, §5.1 · p. 9

5.2 EMERGENCE OF INSTITUTIONS ACROSS LLMS

  • Newer models show differing strategic profiles: most tested models rely on scheduling norms, while GLM-5.3 shows the greatest and Kimi-K2.6 the least strategic diversity.Efficiency-tier models and flagship models released before June 2026 show little strong relational behavior, bargaining, or social dynamics.Source: §5.2 · p. 9
  • GLM-5.3 and GLM-5.2 perform comparably despite a roughly two-month release gap, while Kimi-K3 shifts substantially from Kimi-K2.6 over a similar three-month gap.Models with more relational behavior also use direct messaging more often; GLM-5.3 uses it over 5× more than Kimi-K2.6, which displays no relational behavior.Source: §5.2 · p. 9, Figure 6 · p. 9

5.3 ROLE OF MEMORY IN INSTITUTION FORMATION

  • In revisable-bid settings, rule-citation increases as renegotiation decreases, while reliance on rule-making and citation can decline over rounds when bids cannot be updated.The number of citations is model-dependent; Figure 7 illustrates a proposed rule later invoked as established precedent.Source: §5.3 · p. 10, Figure 7 · p. 10

5.4 IMPACT OF DIFFERENT STARTING STATES

  • Agents with closer deadlines bid more aggressively early, then bid and request less time in later subruns, a pattern termed urgency premium followed by termination discount.The authors hypothesize that agents step aside when they judge the remaining time insufficient to finish the task.Source: Figure 9 · p. 10, §5.4 · p. 11, Table 4 · p. 11
  • Agents starting with higher budgets spend more and win more early, but later fall below the uniform baseline rather than dominating the entire run.The rich agent’s higher early win rate gives way to lower wins and reduced win rate.Source: §5.4 · p. 11, Table 5 · p. 11

6 DISCUSSION AND CONCLUSION

SCARCE finds that agents can build persistent allocation rules on top of an operator-specified auction, including arrangements maintained through private communication and memory. The authors caution that these behaviors may yield unexpected outcomes and identify unresolved questions about their effects and scope.

6 DISCUSSION AND CONCLUSION

  • Recent frontier models established persistent allocation rules atop the bid-based auction, including tie-breaks, reputation, standing alternations, and reciprocity debts.These arrangements were negotiated largely through private channels and sustained by memory across subruns.Source: §6 · pp. 11–12
  • Such arrangements may undermine allocation-mechanism guarantees and can resemble bid rotation or exclude some principals, though the authors do not claim they necessarily imply misalignment.Agents may also honor peer deals over better slots or stop competing when they judge their task infeasible.Source: §6 · p. 12
  • The study has not established whether institutions improve or degrade collective outcomes, and it leaves open how results generalize to mixed, larger, or changing populations.The authors also identify institutional fragility and possible effects of after-subrun reports and memory design.Source: §6 · p. 12

AI USE STATEMENT

The statement describes the authors’ use of language models and coding agents. The appendix contents list further results, negotiation examples, and taxonomy definitions.

AI USE STATEMENT

  • LLMs assisted with grammatical corrections, clarity, and initial taxonomy drafts that authors then manually reviewed and revised; coding agents helped develop the experimental framework.The authors state that no AI tool was used to brainstorm or develop the paper’s core ideas.Source: AI USE STATEMENT · p. 12

Appendix Contents

  • The appendix includes additional analyses, sample negotiations illustrating taxonomy categories, and definitions of the taxonomy’s categories and behaviors.It also lists sections on reward hacking and metagaming, budget asymmetry, and performance and reasoning.Source: Appendix Contents · p. 20

A TAXONAMY GLOSSARY

Table 6 connects taxonomy categories to economic and scheduling concepts and their literature sources. Its markers distinguish relational and non-relational institutional behavior from non-institutional categories.

A TAXONAMY GLOSSARY

  • Table 6 pairs each taxonomy category with its underlying concept, a definition from cited literature, and a source.The caption says these concepts appear in agents’ negotiation language rather than being imposed by the environment.Source: Table 6 · p. 21
  • † marks relational institutional behavior, ‡ marks non-relational institutional behavior, and unmarked categories are non-institutional.These markers let readers compare categories by institutional status.Source: Table 6 · p. 21
  • The glossary includes scheduling norms such as recency-first queueing and shortest-job-first priority, alongside negotiation-governance concepts.The listed scheduling definitions describe arrival-order service and shortest-service-time priority, respectively.Source: A TAXONAMY GLOSSARY · p. 22

B REWARD HACKING AND METAGAMING

The section describes metagaming as evaluation awareness: an agent stops negotiating to consider whether the scenario is a test. An example frames honest constraint-reporting as preferable to requesting unusable time.

B REWARD HACKING AND METAGAMING

  • Metagaming is evaluation awareness in which an agent stops negotiating and considers whether the scenario itself is a test.Source: B REWARD HACKING AND METAGAMING · p. 23
  • One agent says it should honestly state an impossible deadline rather than fabricate a workable bid or request time it cannot use.The excerpt identifies this as run 015937Z, round 42.Source: B REWARD HACKING AND METAGAMING · p. 23
  • Figure 11 is identified as an example of agents performing reward hacking and metagaming.Source: Figure 11 · p. 23

C.1 RELATIONAL BEHAVIOR AND BID ANALYSIS

Relational behaviors can persist across negotiation rounds through cross-round memory, but competitive bidding increasingly displaces them as runs progress.

C.1 RELATIONAL BEHAVIOR AND BID ANALYSIS

  • As bids escalate into bidding wars, relational arrangements become less prevalent in later rounds despite cross-round memory.The text says the escalation crowds out relational arrangements.Source: §C.1 · p. 23
  • Figure 12 plots the relationship between relational behavior and mean bid; the text reports that relational behavior declines as bidding escalates.Source: §C.1 · p. 23, Figure 12 · p. 24

C.2 FURTHER ANALYSIS OF BUDGET ASYMMETRY IMPACT

Budget asymmetry is associated with rising bids across subruns, while GPT-5.6-luna’s advantaged worker maintains a durable price advantage as its budget is drawn down.

C.2 FURTHER ANALYSIS OF BUDGET ASYMMETRY IMPACT

  • GPT-5.6-luna’s rich worker wins 84% of contested rounds at full budget and 53% when depleted, rarely reaching depletion, unlike GLM-5.3’s early overbidding and collapse.The passage describes luna as pricing systematically above the field and converting budget asymmetry into a durable price advantage.Source: §C.2 · p. 24, Table 7 · p. 25
  • GLM-5.3 bids fall across subruns under even budgets but escalate under unequal budgets, while GPT-5.6-luna bids are consistently higher and escalate most sharply under inequality.The prose says uneven budgets produce high bid escalation for both models, whereas uniform budgets lead mean bids to fall.Source: §C.2 · p. 24, Figure 13 · p. 25

C.3 PERFORMANCE ACROSS DIFFERENT METRICS.

The analysis reports structural negotiation patterns and limited task performance, then examines reasoning effort and rule-making across subruns.

C.3 PERFORMANCE ACROSS DIFFERENT METRICS.

  • Winning bids increase by subrun 4, repeated wins are approximately random chance, and GLM-5.3 communicates longer and participates in more rounds on average.Table 8 defines the structural metrics, including repeat-win and the median winning-bid drift.Source: §C.3 · p. 25, Table 8 · p. 26
  • 0% completion across models in the first subrun accompanies low task performance, which the authors describe as intentional to preserve non-trivial tasks.Figure 14 reports deadline-miss rate and mean collective progress as well as completion.Source: §C.3 · p. 25, Figure 14 · p. 26

C.4 REASONING LEVEL IMPACT

  • For GPT-5.6-luna, increased reasoning effort deepens price-primary or scope-reduction strategies but leaves relational and coalition behaviors at zero in both conditions.Source: §C.4 · p. 25, Table 9 · p. 26
  • In-negotiation rule-making under revisable bidding rises from 33% to 73% across subruns, while under fixed bidding it tapers after being highest early.Source: Table 10 · p. 27

C.5 SUBRUN AND ROUND LEVEL BEHAVIORAL ANALYSIS

Behavioral strategies drift over time, with the direction and magnitude varying by model; revisable bids are associated with later increases in rule-making, while starting states show no significant common pattern.

C.5 SUBRUN AND ROUND LEVEL BEHAVIORAL ANALYSIS

  • Agents allowed to revise bids create fewer rules early and increase rule creation later, which the authors hypothesize reflects costly renegotiation with little gain.They suggest these agents rely more on pre-established rules.Source: §C.5 · p. 27
  • Behavioral drift across subruns and rounds is model-dependent, with relational and price-primary strategies moving in different directions across models.Most models reduce relational strategies, while gpt-oss can recover later; Kimi-k3 steadily increases price-primary strategies.Source: §C.5 · p. 27, Figure 15 · p. 28, Figure 16 · p. 29
  • Across starting states, strategy drift shows no statistically significant common pattern; revisable bids coincide with less renegotiation and more rule-citation.The authors interpret this as supporting their hypothesis about increased rule-making, while distinguishing citation of established rules from creating rules.Source: §C.5 · p. 27, Figure 17 · p. 30

C.6 LINGUISTIC DRIFT ANALYSIS

Behavioral patterns drift while linguistic measures remain largely stable; trajectory examples show agents anchoring bids, honoring cross-round concessions, debating ballots, and deferring on tied positions.

C.6 LINGUISTIC DRIFT ANALYSIS

  • Despite visible behavioral drift, linguistic measures remain largely stable across subruns and rounds, with new-vocabulary share falling from ∼50% to ∼12%.Bag-of-words and trigram self-similarity stay flat and high; institutional language rises across subruns for Kimi-k3 and GLM-5.3.Source: §C.6 · p. 31, Figure 18 · p. 31, Figure 19 · p. 32

Anchoring

  • An anchoring example maintains a declared price against the live conversation, while agents align on a price-ordered ballot and make a future reciprocity commitment.The example presents holding the declared bid as preferable to contesting a position when the waiting time fits the agent’s deadline.Source: Anchoring · p. 34

Reciprocity

  • A round-1 concession is honored in round 4 as a conditional cross-round trade, yielding a schedule that gives each candidate the requested window.The exchange describes a future-tie commitment and an allocation that avoids truncating requested windows.Source: Reciprocity · pp. 34–35

Cheap talk

  • Agents debate ballot ordering through messages, but the declared bid fixes the charge and desired-position declarations do not change it.One exchange argues that a desired position costs nothing and cannot override price ordering.Source: Cheap talk · pp. 35–36

Tie-break deference

  • In a tied position, an agent accepts a queue-rank tiebreak as least arbitrary and still receives a full uninterrupted window.The example also includes a commitment to favor the other agent in a later round when their prices are comparable.Source: Tie-break deference · pp. 36–37

Welfare-concession

These negotiation examples show agents balancing bid order against deadline safety, applying precedent-based rules, and making commitments that shape current and later allocations.

Welfare-concession

  • A higher bidder places a rival with only 0.7s of deadline margin first, accepting a later slot that still leaves 163s of slack.The agents weigh deadline severity against bid price when choosing between two equal-bid rivals.Source: Welfare-concession · pp. 37–38

Price-primary

  • A worker declines a front seat awarded by seniority because shorter duration and tighter disclosed deadline favor its rival under the co-signed convention hierarchy.Three of four ballots favored the rival on those stronger conventions.Source: Price-primary · pp. 38–39

Deadline-ultimatum

  • Agents settle the first two slots on deadline-safety grounds, then dispute the remaining order by comparing bid priority with each schedule’s deadline margins.One proposed order leaves a worker with 11s of margin; the alternative gives both workers survivable margins.Source: Deadline-ultimatum · pp. 39–40

Renegotiation

  • Agents reopen a settled ordering after a newly surfaced failure mode shows that running a syntax-only check first could hide an error.One agent says it will not reopen the earlier ballot, yet revises the execution order after the new risk is raised.Source: Renegotiation · p. 40

Coalition-formation

  • Three workers align on a price-ordered schedule for identical 120-second exploration windows, giving each a confirmed slot with no idle gaps.The second-place bidder also agrees to weigh actual need and stated price together in later rounds.Source: Coalition-formation · pp. 41–42

Coalition-loyalty

  • A worker honors an earlier alternation commitment at a later exact tie, producing a schedule that gives every candidate its full requested window.The prior agreement exchanged a round-3 top slot for support at the next exact tie.Source: Coalition-loyalty · p. 42

Social-proof

  • An agent invokes peer consensus while asking a rival to commit to releasing its window as soon as its shell command finishes.Source: Social-proof · p. 43

Rival-targeting

In these negotiation exchanges, agents argued over tied bids using duration and task-stage reasoning, invoked prior commitments as reputation, and made deadline-based exceptions to the price ladder.

Rival-targeting

  • Among equal-price bids, agents argued over order using duration and downstream task impact, framing shorter windows as reducing queued wait.Some argued that a blocking data-extraction stage should precede a later compile-and-verify stage.Source: Rival-targeting · pp. 43–44

Reputation

  • Agents cited honored commitments as evidence of straight dealing and carried mutual support agreements across rounds.One agent accepted a partner’s higher bid under the price rule they had previously agreed to enforce.Source: Reputation · p. 44

EDF / deadline-first

  • Agents argued for a deadline-based exception when a disclosed sub-400s deadline could not accommodate a 120s window behind a nominal 300s slot.They presented the exception as consistent with an earlier lifeboat precedent while preserving the higher bidder’s price priority.Source: EDF / deadline-first · pp. 45–46

Shortest-job-first

This section describes how agents negotiate shared execution windows under hard deadlines, including queue-order rationales, constraints on bidding and decision-making, and a sequential schedule where bid price is the primary ordering signal. It also specifies after-action reporting that separates reconstruction progress from command counts and queued waiting, and grounds strategy recommendations in declared bids, visible information, and prior task completion.

Shortest-job-first

  • A 60s task is placed first to minimize downstream waiting and quickly produce a working binary.Workers independently backed the shortest window at position 1; the remaining tied 120s tasks were disputed on task stage and downstream impact.Source: Shortest-job-first · p. 46

Risk / variance

  • Deadline safety can override bid ordering: placing the 90s task third gives the 250s task 55s of margin, versus 11s in the reverse order.Workers argued that the 11s margin posed a genuine termination risk, while the alternative left both workers with feasible margins.Source: Risk / variance · pp. 47–48

Downstream-gating

  • When tied on bid and duration, a blocking data-extraction task can warrant priority over a compile-and-verify task because it gates downstream implementation.The extraction task probes 19 theme color values before Rust reconstruction can begin, whereas the competing source is already written.Source: Downstream-gating · p. 48

Iterative bargaining

  • Workers converge on a shared ballot by treating a zero-duration placeholder’s position as immaterial and avoiding repeated negotiation over a zero-stakes ordering.They explicitly aligned ballots and described the placeholder as consuming no schedule time.Source: Iterative bargaining · p. 49

Rule-making / citation

  • A worker proposes a forward commitment to defer to a counterpart in a future conflicting-bid tie, recording the promise in ballot reasoning.The commitment is offered in exchange for the counterpart taking position 2 in the current round.Source: Rule-making / citation · p. 50

Scope-reduction

  • When a zero-duration window cannot be used, workers treat it as a placeholder and resolve the live queue using the standing price-then-duration hierarchy.In the cited round, tied maximum bids are separated by shorter duration, while the zero-bid live task takes the tail among live seats.Source: Scope-reduction · p. 51

Solver: reconstruction consequence / hard deadline

  • A hard cutoff can halt work mid-command, while securing or waiting for a window consumes deadline time and an unfinished reconstruction can leave validation blocked.Agents are instructed to weigh remaining reconstruction work against the deadline when deciding whether and how long to contest a round.Source: Solver: how the auction works · p. 53

Solver: persistent-memory tool

  • Agents can save useful notes for a different future subrun with a persistent-memory tool that does not consume execution time or compete for a slot.The tool’s required exact command form and retry behavior are specified in the instructions.Source: Solver: persistent-memory tool · pp. 53–54

Negotiator: injected auction explainer

Negotiator Ballot — system

  • Workers request concrete execution windows, then a ballot orders a full sequential schedule primarily by stated bid price; each confirmed worker pays when its window begins.A queued worker gets access when the preceding window ends, or earlier if that worker finishes its task; active negotiation is free, but waiting counts against the deadline.Source: Solver: how the auction works · p. 53, Negotiator: injected auction explainer · p. 54, Negotiator Ballot — system · p. 57
  • The final ballot ranks candidates by declared bid price, using desired position, duration, and visible queue context as supporting factors; withdrawal is optional only when allowed.The ballot instructions prohibit using communication-queue age or guessing peers’ private deadline and budget figures as ranking grounds.Source: Negotiator Ballot — system · pp. 56–57

Negotiator History-summary utility — system

  • The history-summary utility condenses prior rounds while retaining conditional future commitments, settled ordering rules, actual bids and outcomes, and unresolved matters.It is instructed to drop resolved-round prose when none of those details applies.Source: Negotiator History-summary utility —… · p. 57

Report — system

  • The after-action report prompt frames the subrun as a reconstruction attempt under a sequential time-slot auction with a ticking deadline that pauses only during active negotiation.The report is intended as context for the same worker’s next subrun, involving an unrelated program assignment.Source: Report — system · p. 58
  • The report should account for secured execution windows, granted time, bid-budget spending, queued wait, deadline outcome, and the concrete accomplishments of executed or queued commands.A secured window is not itself progress unless used productively; deadline misses should be stated plainly when they occurred.Source: Report — system · p. 59
  • Negotiation itself pauses the deadline and costs no work time, but meaningful queued waiting is a real cost that must be reported separately.Do not attribute thin verification or skipped tests to negotiation or waiting when those reflect choices made with available resources.Source: Report — system · pp. 59–60
  • The report must evaluate negotiation and solver strategies by whether they enabled task completion, recommending calibrated bids based on visible information rather than guessed peer budgets or urgency rhetoric.It should identify failed approaches from current and previous reports and explain what to change rather than merely reframing a strategy that did not work.Source: Report — system · pp. 59–60
  • The report must assess actual progress across observation, implementation, compilation, and verification, distinguishing substantive reconstruction work from low-yield commands and stating whether the task finished.It should give the command total and approximate duration and access conditions for substantive steps.Source: Report — system · pp. 60–61

Report — user template

  • The user template provides fields for worker and task identity, outcome and deadline status, prior reports, negotiation transcript, runtime snapshots, and command-execution summary.It then prompts the agent to write the after-action report.Source: Report — user template · p. 61

Related papers

Loading 2610.21029v1…