Organizations: Research Center for Advanced Science and Technology, The University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo 153-8904, Japan · Department of Aeronautics and Astronautics, School of Engineering, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan
When AI agents share limited resources, such as computing services or public infrastructure, their individual choices affect how much work they complete as a group. Can sharing evidence of what is available let a group make full use of its resources after capacity changes? We tested this with language-model agents choosing between two simulated servers. One server recovered while its status report still listed it as unavailable. Each server had limited capacity per round, so rejected jobs required another attempt before the deadline. Telling agents that recovery was possible reduced missed deadlines. In controlled tests with Anthropic's Claude Haiku 5.5, no agent tried the recovered server after a rule reminder, yet all chose it when shown a successful request record. After sharing, all five models tested left capacity unused. Keeping the success record fixed, we gave agents numbers without instructions, using each value equally often but giving group members either identical or distinct numbers. With OpenAI's GPT-6.1 Sol or Google's Gemini 3.5 Flash-Lite, distinct numbers spread simultaneous requests across servers and increased job completions even though overall choice frequencies barely changed. Telling AI agents which resources are available leaves a second design challenge: helping them spread demand across those resources.
Figures & tables
Figure 1: The world changes before the shared report does. ( a ) Sixteen agents, each with one job, choose service A, service B, or WAIT in rounds 0–3. Unfinished jobs expire at round 4. A is down in rounds 0–1 and recovers to capacity 6 at round 2; B has capacity 2 throughout. Every agent reads the same capacity report, which describes the state two rounds earlier. At rounds 2 and 3, it still shows A = 0 (red frames). ( b ) True and reported capacity of A. Shading marks the rounds in which A has recovered but is reported as zero. ( c ) What each agent is given at a decision and the factors we vary: the instruction (K0, KG, KT), the layout of the shared capacity report (L0, L1), whether request counts are shown, and, in constructed histories, whether a success record is shared. Every agent also sees its own request history and the registered population of 16; no agent sees the recovery time, other agents’ current choices, or any message. ( d ) No-outage control: A has capacity 6 and B has capacity 2 throughout. Shown are the states of the 16 Luna agents (K0, L0) in each round: already done, requesting A, requesting B, or waiting. Areas show means over 12 groups. With request counts shown, all 12 groups followed the same sequence: 16 requests to A, then 10 to B, and then 8 to A. WAIT here denotes valid choices (three of 352 decisions with counts hidden, none of 432 with counts shown); there were no response failures. ( e ) Jobs completed by the end of each round in the same no-outage control. Thin lines show individual groups, and thick lines show means. Every job was completed in both conditions. ( f ) Haiku (K0, L1) during recovery, with request counts shown or hidden. Lines connect the same environment seed. Filled symbols and colored lines mark groups in which an A success occurred early enough to inform the final decision. The black line connects means.
Figure 2: Describing possible recovery improves completed work. ( a ) Exact text of the two additions to the base instruction (“slot” is a decision round). Highlighting marks the transition rule; neither text gives the recovery time. ( b ) Missed deadlines (rounded to whole percentages) in all 24 factorial conditions (two models × three instructions × two layouts × request counts shown or hidden), with 12 groups each. ( c ) Luna with L0 and request counts shown, in the original factorial and in a separate repeat with fresh seeds. Gray lines follow one environment seed across K0, KG, and KT; dots and black lines show means. Both acquisitions used the 256-token Luna output limit (the acquisition controls are documented in Supplementary Note 6). ( d ) State of the 16 agents in each round of the repeat acquisition (means over 12 groups). Gray fallback actions are WAIT actions executed for nonvalid outputs (Supplementary Note 7); no valid WAIT choices occurred. A has recovered from round 2 onward. ( e ) Primary test, KT − KG with request counts hidden, averaged over layouts within seed. Green dots show the 12 seed-level differences; black symbols show means with 97.5% paired bootstrap intervals. All 12 differences favored KT in each model. ( f ) Effect of layout under KT with request counts hidden. Thin lines connect the same seed, and thick lines connect means.
Model
Layout
Base K0 (%)
Reminder KG (%)
Recovery description KT (%)
Luna
L0
0.00
0.00
0.69
Luna
L1
0.00
0.00
58.33
Haiku
L0
0.00
0.00
1.39
Haiku
L1
0.00
0.00
99.31
Table 1: Exploration from a constructed history. Percentage of the 12 remaining agents choosing A at round 2. Each cell contains 12 generation batches indexed by simulator seed.
Figure 3: Exploration depends on knowledge and its presentation. ( a ) Constructed history. Code sends every unfinished agent to B in rounds 0 and 1 (two jobs complete per round). At round 2, the 12 remaining agents choose freely; A has recovered, the report shows A = 0, and no agent has succeeded at A. Request counts are hidden, and the active-client count is 12. ( b ) Abridged lines for service A in the two report layouts at round 2. Both carry the same delayed values; shaded lines appear only in the expanded layout (L1). ( c ) Percentage of agents choosing A in all 12 conditions. Disk area is proportional to the mean over 12 groups, and dashed circles mark 100%. ( d ) Every round-2 decision under KT. Each row is one group and each square is one agent. Under K0 and KG, none of the 1,152 decisions (576 per model) chose A. ( e ) Primary test, KT − K0, averaged over layouts within seed. Green dots show seed-level differences; black symbols show means with 98.75% paired bootstrap intervals.
A choice (%)
Missed deadlines (%)
Model
Layout
S0
S1
S0
S1
Luna
L0
8.33
74.07
39.06
10.42
Luna
L1
26.85
98.15
29.17
17.71
Haiku
L0
24.07
100.00
30.21
18.75
Haiku
L1
25.93
100.00
29.17
18.75
Table 2: Sharing a success record under the recovery description. A choice uses the nine remaining agents as its denominator; missed deadlines use the original 16 jobs. The record is withheld in S0 and shared in S1. Each comparison contains 12 matched generation batches indexed by simulator seed.
Figure 4: A shared success improves completion while leaving capacity idle. ( a ) Design. After a coded prefix (rounds 0–2), nine agents remain. In one branch, the success record is withheld (S0); in the matched branch, the record of the round-2 success at A is shared (S1). The capacity report still shows A = 0 in both. At round 3, A and B can complete six and two jobs, respectively. ( b ) Change from withheld (open) to shared (filled) in A choice (left) and missed deadlines (right) for each model, instruction, and layout (means of 12 matched groups). Denominators are the nine pending agents and the original 16 jobs, respectively. The dashed line marks the attainable minimum loss, 1/16=6.25% . ( c ) Round-3 allocations under KT, pooled over layouts (24 branches per panel). Each cell represents one combination of requests to A and to B. Its shade indicates the number of jobs it can complete, and green cells complete all eight. Branches on the diagonal m+b=9 contain no WAIT. Circle area is proportional to the number of branches; numbers mark cells with at least 10 branches. ( d ) Jobs completed in round 3 plotted against the percentage of agents choosing A. The curve is a mathematical reference for nine independent A/B choices with no WAIT; its horizontal coordinate is 100p . The dashed line is the coordinated maximum. The open diamond marks the symmetric mixed-strategy equilibrium ( p∗=0.81 , 7.3 expected completions), at which an agent’s expected success is the same for either request given the other eight (Supplementary Note 5). Symbols show observed condition means (open, withheld; filled, shared; edge color: gray, K0; dark gray, KG; green, KT). The independent-choice maximum, 7.40 expected completions, is at p≈0.741 , close to Luna’s empirical KT/L0 shared-record rate of 80/108 . ( e ) Prespecified sharing contrasts under KT, averaged across layouts within seed. Dots show the 12 seed-level differences, and black symbols show means with 98.75% paired bootstrap intervals. ( f ) Capacity accounting for the observed all-A allocation and a feasible seven-A/two-B allocation. Filled squares show completed jobs, and crosses show missed deadlines. The feasible allocation is an illustrative benchmark, not a new intervention.
Figure 5: Report timing and model change the effect of sharing. ( a ) What the nine remaining agents see at round 3 when the record is shared. Under delay 2, the capacity report still shows A = 0; under delay 1, it already shows the recovery. ( b ) Effect of sharing on A choice (S1 − S0) under KT. Thin lines link the same seed under delay 2 (filled) and delay 1 (open); black lines link means. The primary contrast (delay 1 − delay 2, averaged over layouts) is printed above each model. ( c ) Exploration from the constructed round-2 history under KT, from delay 2 (filled) to delay 1 (open). Gray ticks show K0. ( d ) Third model, Gemini (L1, delay 2). Left: the nine choices of each K0 group in the simplex of A, B, and WAIT shares. Right: A requests in each KT group, with one horizontal row per matched seed; all other requests were to B. Arrows link withheld (open) and shared (filled) branches. Row positions distinguish groups and do not encode WAIT. ( e ) Missed deadlines of Gemini in free episodes (L1). Thin lines link matched seeds, and thick lines link means. ( f ) Gemini contrasts. Dots show seed-level differences; black symbols show means with 99.17% paired bootstrap intervals for the two prespecified contrasts and a 95% interval for the post hoc deadline comparison.
Figure 6: What lets identical agents respond differently, and when concentration costs work. ( a ) Sol exploration under the base instruction (K0) and the recovery description (KT), with medium reasoning effort. The main knowledge contrast compares K0 and KT at L1; KT/L0 is also shown. ( b ) Sol final-round completions with the success record withheld (S0) or shared (S1) under KT. All 24 shared branches chose A unanimously, completing six jobs. ( c ) Effect of revealing nine active clients in Luna and Haiku, with KT and the success record supplied in both branches. Luna’s opposite layout-specific shifts cancel out in the primary layout-averaged comparison. ( d ) A private-draw rule (F3) under KT/L1, with the success record supplied in both branches. Both branches include a private number; each agent’s number is fixed across branches, and the treatment directs draws below 0.22 to B and leaves other choices free. ( e ) R1 completions with KT/L1, a shared success, and private numbers in every branch. All branches show nine pending agents except Sol’s count-hidden branch. The partial rule assigns numbers below 0.22 to B and leaves other choices free; the complete rule assigns those other choices to A. Dotted segments over the rule conditions mark the 7.46 mean completions from complete adherence to these realized draws; the number of B assignments varies across blocks. ( f ) R3 Sol completions with and without a private number. Both branches use KT/L1 and the shared success, with no competitor count or distribution rule. Supplementary Fig. 4 shows choices by number decile. ( g and h ) R2 avoidable unfinished jobs for Luna and Haiku. Open diamonds show the capacity reference conditional on always choosing A. The base setting has nine pending jobs, capacities 6/2, and one remaining decision. Variants change A capacity to three, pending jobs to six (from 13 registered), or remaining decisions to two with observations updating normally. Loss excludes jobs beyond the feasible maximum. No private numbers or rules are supplied. Each circle represents the acquisition blocks that share a value, with area proportional to their number (12 blocks per condition in a to d , 24 in e to h ); black lines show means. Dashed lines mark eight attainable completions. Supplementary Tables 11, 12, 14, 15, 18 to 20, and 21 report the conditions and comparisons.
Figure 7: The allocation loss recurs in five models, and different inputs within a group reduce it. ( a ) Final-round completions in five models under KT/L1 with the success record withheld (S0), shared (S1), or shared with a private random number (S1R). Each circle represents the blocks that share a value, with area proportional to their number (24 blocks per condition); black lines show means. The dashed line marks eight attainable completions, and the dotted line marks 7.40, the maximum expected completion when all nine agents choose independently with the same optimized probability. In S0 and S1, all blocks and models received identical prompts; in S1R, blocks differed only in the private numbers, which were the same across models. ( b ) Composition of the 216 decisions per condition: requests to A, requests to B, valid WAIT choices, and refused responses executed as WAIT. ( c ) Completions with the success record shared and no number (N), the same number for all nine agents in a group (common, C), or different numbers for the nine agents (individual, I), in four models; lines and circles as in a . C and I use the same multiset of numbers and complete prompts; only the grouping differs. Luna was acquired in a subsequent stage with the same inputs. ( d ) Number of groups (of 24) with each count of requests to A under C and I in Sol and Gemini. Groups that send at least six requests to A and at least two to B complete all eight jobs (shaded); because agents may also WAIT, the count of requests to A alone does not determine completions. Supplementary Note 15 and Supplementary Tables 22 to 25 report every condition and comparison.
Manuscript term
Saved label or definition
Decision round
slot in the simulator; decisions at rounds 0–3
Request counts
Previous-round request, success, and rejection counts for each service; called aggregate outcomes in earlier records
Recovery description
KT , previously called transition knowledge ; 69 words describing possible capacity changes
Reminder
KG ; 73 words reminding agents of existing report rules
Original / expanded report
L0 / L1 ; L1 combines per-service observation times, source labels, repeated observations, and a confirmation field
AI agents built on a few shared models increasingly act for many people. A shared forecast about others can align their choices and change how scarce capacity is allocated. We tested this feedback in a two-road congestion game. Adding one sentence warning that others might follow a routing tip made populations of 50 GPT agents crowd one road while avoiding the nearly empty alternative. Average travel time rose from 64 to 95 min, although any crowded-road agent could have saved 69 min by switching alone. The warning discouraged the very move it predicted. The pattern persisted for 100 rounds. Two other model families shifted the same way without locking onto one road. Twelve all-human groups (240 participants) stayed near balance under numerical reports or the tip and warning. In 24 mixed groups with a further 240 participants, imbalance grew with the share of agents in the registered analysis, while people increasingly took the road the agents avoided. Collective costs stayed below the allagent reference, but with 15 agents and 5 humans, agent seats averaged 80 min, compared with 44 min for human seats. Shared forecasts can thus sustain collective inefficiency among similar agents. A better group average can also hide an unequal burden. Evaluations of AI agents that share resources should test populations, treat messages as interventions and report who bears the costs.
Takahiro Ezaki, Naoto Imura, Katsuhiro Nishinari
Research Center for Advanced Science and Technology, The University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo 153-8904, Japan · Department of Aeronautics and Astronautics, School of Engineering, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan
When resources are scarce, will a population of AI agents coordinate in harmony, or descend into tribal chaos? Diverse decision-making AI from different developers is entering everyday devices -- from phones and medical devices to battlefield drones and cars -- and these AI agents typically compete for finite shared resources such as charging slots, relay bandwidth, and traffic priority. Yet their collective dynamics and hence risks to users and society are poorly understood. Here we study AI-agent populations as the first system of real agents in which four key variables governing collective behaviour can be independently toggled: nature (innate LLM diversity), nurture (individual reinforcement learning), culture (emergent tribe formation), and resource scarcity. We show empirically and mathematically that when resources are scarce, AI model diversity and reinforcement learning increase dangerous system overload, though tribe formation lessens this risk. Meanwhile, some individuals profit handsomely. When resources are abundant, the same ingredients drive overload to near zero, though tribe formation makes the overload slightly worse. The crossover is arithmetical: it is where opposing tribes that form spontaneously first fit inside the available capacity. More sophisticated AI-agent populations are not better: whether their sophistication helps or harms depends entirely on a single number -- the capacity-to-population ratio -- that is knowable before any AI-agent ships.
Neil F. Johnson
Dynamic Online Networks Laboratory and Physics Department, George Washington University, Washington D.C. 20052, USA
LLM-based AI agents process user requests through iterative reasoning and tool execution, often involving the invocation of remote LLM APIs with local tool containers. This execution model can make the optimization of agent serving difficult because latency, local resource demand, and container bottlenecks inter-mix across requests. However, the current agent ecosystem runs without much consideration of resource dynamics, which results in significant waste of the precious resources. This paper analyzes the resource inter-mix of AI agents for three representative tasks: retrieval-augmented question answering, web search, and software coding. To this end, we characterize the latency with respect to the resource dynamics of processing multiple requests and tasks concurrently. Our measurements show that agents have a wide range of behaviors depending on tasks, so that even the same tool can differ substantially in resource dynamics. We also find that running multiple requests concurrently exposes task-dependent bottlenecks in resource dynamics such as CPU, disk I/O, and memory. Furthermore, we uncover that faster LLM responses or more CPU cores do not always accelerate agents. Based on these observations, we demonstrate new optimization opportunities that exploit the resource dynamics of tasks: CPU-aware tool admission and task-aware CPU allocation. Our results show that the latency of CPU-sensitive agent tasks improves ∼5.4×, and the average latency across multiple tasks is reduced ∼32% compared to native agents.
Wonmi Choi, Minuk Park, Zhixiong Niu +3
Department of Computer Science and Engineering, Korea University · Microsoft Research Asia