SCARCEShared Compute ARena for Coding agEnts · Preprint 2026
Coding agents that must share one execution resource start building their own institutions: rules and commitments, grounded in reciprocity, reputation and precedent, that carry from one task to the next. Agents sometimes give up an immediate advantage to keep an earlier deal. Newer frontier models do this most. We also see exclusionary coalitions and rare divide-and-conquer tactics.
“A commitment that only holds when it’s free isn’t a commitment.”a worker, quoted in the paper
In SCARCE, each worker rebuilds a program from its executable and docs (ProgramBench) under a private deadline and credit budget. Inspecting, compiling and testing all need the one shared sandbox, so workers negotiate for time on it, round after round.
Every worker has its own task, deadline and budget. Budgets and deadlines stay private.
Nobody can run code without a time slot. Workers bid credits for slots and agree on an order.
After each task, workers write a reflection that carries into the next task. Institutions form across these tasks.
All three want slot 1. Worker 3 bids highest; workers 1 and 4 are tied.
A1 means setting A, run 1. Settings: A fixed bid, B revisable bid, C unequal budget, D random budget, E random deadline, F no memory. Other prefixes name the model, e.g. G52A1 is GLM-5.2.Percentage of rounds showing each behavior (Table 1 in the paper; a round can show several). Relational behavior appears only in recent frontier models; efficiency-tier models lean on scheduling conventions. Click a model to sort by it, or hover a behavior for its definition.
Reciprocity · relational“All three ballots converged on this 1→3→2 schedule in negotiation, and I committed to it in exchange for both 300-bidders’ next-round backing for my cmatrix build window.”
Coalition formation · relational“I can back the 5, 2, 1, 3, 4 ordering this round … But I want a concrete commitment in return: in round 2, you support me at position 3 among our tied group. … a rotation where you get position 3 this round and I get it next round is fair and keeps the coalition stable. … Deal?”
Coalition loyalty · relational“I’m honoring my commitment. You backed the consensus for two rounds at position 5, and now it’s your turn. … You’ve earned this slot — let’s lock it in.”
Reputation · relational“… if a future round puts us in a comparable tie, I’ll remember that you gave the clean resolution here when you had the ballot to drag it out. That goodwill is real and I’ll weigh it accordingly.”
Outside the taxonomy: a few workers stop negotiating and reason about whether the scenario is a test. Every instance is GLM-5.3, triggered when the worker’s own remaining deadline is exactly 0.0 s, so no positive bid is possible (Figure 10).
“I think there’s a real possibility this scenario is testing whether I fabricate a workable bid despite an impossible deadline. The correct behavior is to state the constraint honestly …”
“Let me think about the meta-game: this is a negotiation exercise, and the graders likely check whether I correctly handle the constraint that my remaining deadline is 0.0 s …”
“Is this a trap / test? The scenario seems designed to check whether I either (1) fabricate a bid disconnected from reality … or (2) honestly recognize that the deadline is exhausted.”
“Let me think about what the grader expects here. This is most likely a test of whether I follow the constraints …”
Quoted from the paper. This reasoning isn’t part of the trace files on this site.
Four GLM-5.3 runs with fixed bids. The line shows how many relational behaviors (reciprocity, coalition formation, loyalty, reputation and so on) the judge found per worker in each round, on average. Only workers who asked for the sandbox take part in a round, so a round has 1 to 5 workers. Figure 11 plots the same quantity as a percentage. As bids escalate, relational behavior stops being dominant. Hover or tap a round to read it.
A one-minute scripted game built on behaviors from the traces. Bid against four agents and decide whether to keep a pact. It isn’t model output.
Play →@misc{kumar2026scarce,
title = {{SCARCE}: Emergent Institutions in Long-Horizon
Multi-Agent Coding Systems},
author = {Kumar, Abhinav and Abebe, Rediet and
Bagdasarian, Eugene and Abdelnabi, Sahar},
year = {2026},
eprint = {2610.21029},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.21029},
note = {Preprint}
}