Experiments with remote decision-making agents depend on more than the policy: the world may change during inference, the simulator determines which actions are available, and births and immigration alter who can act. We present Jevolution, an instrumented spatial predator-prey framework that makes these dependencies explicit. Its reference implementation separates the evolving world, asynchronous request scheduling, controller adapters, and experiment records. The model combines local observations, energy and resource constraints, predefined signals, and inherited scalar traits. We describe these mechanisms using the ODD protocol and distinguish implemented controls from research questions requiring additional interventions. Free synthetic fixtures exercise scheduling, signaling, and inheritance against the pinned implementation; they validate specified software behavior rather than model intelligence or ecological realism. A completed communication case study illustrates the end-to-end workflow using Jev to control both species. Across 6 complete habitat-seed pairs, the mean signals-on minus signals-off difference in normalized rabbit abundance was -0.0095 (nominal 95% pair-bootstrap interval [-0.0772, 0.0602]). This small study does not resolve the direction of the signal-system effect. The contribution is an inspectable experimental system and its evidence boundaries, not a new policy-learning algorithm or a demonstration of emergent language, adaptive selection, or biological predictive validity.
Figures & tables
Figure 1: Architecture of the pinned implementation. Arrows show information flow, not synchronous execution; world physics continues during requests. The controller receives local observations and engine-offered choices; the engine resolves routes and applies effects. Request identifiers join records across both paths. This schematic is not an empirical result, so no seed count or confidence interval applies.
Figure 2: Jevolution’s interactive interface. Left: run controls and population history. Center: habitat, animals, and signal markers. Right: the selected animal’s decision and state inspector. The image is reproduced unchanged from Figure 1 of the authors’ original manuscript, at 31.59 simulation seconds; it is not one of the twelve 90-s study runs. Visible signals illustrate the interface, not evidence of a communication benefit. Interface and artwork provenance are documented in Section 8 and Appendix K .
Route
Supported operation and boundary
Run configuration
Seed, communication condition, timing settings, and audit runtime options. Predator–prey reset installs a fixed Jev roster: supplied provider, population, and group-delay roster edits are replaced. Arena roster configuration is a separate path.
Injected service
A host may replace choose while keeping the engine and observation builder. Synthetic responses test this boundary; virtual clocks additionally require the white-box harness described in Appendix C .
Source changes
New species, a supported clock port, receiver-only/fixed-trait controls, or policy optimization need additional implementation and validation. Frozen UI preview metadata is not an active experimental control.
Table 1: Actual intervention routes on the frozen source. This is a source-backed capability map, not a usability experiment.
Repeated patterns and spatially matched nulls preserving resource attraction and demographic placement.
Natural selection
Inherited traits, ancestry state, and immigration counters; cohort reconstruction required
Cohort/lineage reproductive success and fixed-trait or inheritance controls separating mutation, drift, immigration, and selection.
Policy learning
Controller adapter boundary
A training loop, learning objective, and evaluation protocol; EPO is not implemented.
Table 2: Research scope of the reference system. These are capability/evidence distinctions, not empirical estimates; no seed count or interval applies. The case study and synthetic checks are reported separately.
Injected delay (ms)
Capacity 1
Capacity 3
0
238 / 238 / 0 / 0
240 / 240 / 0 / 0
100
76 / 74 / 0 / 2
126 / 120 / 0 / 6
500
18 / 16 / 0 / 2
42 / 36 / 0 / 6
1000
10 / 8 / 0 / 2
24 / 18 / 0 / 6
1001
10 / 0 / 8 / 2
24 / 0 / 18 / 6
1500
6 / 0 / 4 / 2
18 / 0 / 12 / 6
Table 3: Synthetic runtime fixtures, not provider throughput: request/applied/late/cancelled counts per execution over 4.137 simulation seconds. Each cell was run for two seeds (42, 314159), each repeated twice; all four executions in a cell had the shown counts. There is no sampling-based interval for these deterministic fixtures. Capacity is per lineage, with three animals in each of two lineages. The acceptance deadline is 1,000 virtual wall ms, distinct from the live study’s 2,500 wall ms.
Figure 3: Communication worked example: all n=6 paired normalized rabbit-AUC effects and their mean. Each pair contains one 90-s run per arm. The nominal 95% percentile-bootstrap interval uses 10,000 whole-pair resamples (seed 20260928); few-pair coverage is unreliable. These are the same recorded runs as the communication-focused draft, not an independent replication.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Coordinates from retained world states for pair 01 (habitat seed 80350226): the shared initial state and the two terminal states at 90 simulation seconds. This is a visualization of recorded data, not a screenshot captured during execution. It shows one of the six pairs ( n=1 illustrated pair, two runs); the first numbered pair was chosen by index during figure preparation, after data collection, rather than by its effect or appearance. The terrain and initial animal coordinates match across arms. Markers show positions; overlapping markers can hide individuals. Food, energy, signals, and ancestry are omitted. No spatial hypothesis or clustering statistic is tested, so no confidence interval is attached; the aggregate paired estimate and uncertainty are reported in Figure 3 .
Environments and components for constructing agent-based models.
Jevolution supplies one specialized population task and a remote-decision execution/recording path; replacement of general ABM tools is not established.
Generative Agents, Project Sid, AgentSociety [ Park et al., 2023 , Altera.AL et al., 2024 , Piao et al., 2025 ]
Agent architectures, real-time social worlds, and simulation/intervention systems.
Real-time agent societies and framework-plus-intervention studies are prior art. This case uses a structured decision interface in a mechanistic population task.
flame / AgentTorch [ Chopra et al., 2024 ]
Differentiable agent-based models for calibration and action learning.
Jevolution does not differentiate through its world or optimize its controller in this study.
PettingZoo / Melting Pot [ Terry et al., 2021 , Leibo et al., 2021 ]
Explicit multi-agent interaction API; evaluation across social situations and partners.
The present service boundary is source-level and domain-specific; API compatibility and equivalent evaluation breadth are not established.
Jevolution
Pinned spatial engine, asynchronous requests, population turnover, signals, inherited traits, and joined records.
Evidence covers the declared software fixtures and one live-model communication intervention. Practical advantage over the systems above remains an open evaluation question.
Appendix
Table 4: Documented system emphases and this paper’s narrower reference task. These are literature/source descriptions, not measured comparative scores.
Pair
Habitat seed
First arm
1
80350226
Signals on
2
1766849094
Signals on
3
1976771215
Signals off
4
865173864
Signals on
5
1328634472
Signals on
6
415864476
Signals on
Appendix
Table 5: Frozen main-study seed blocks and order. Six pairs were planned, with one attempted run per arm and no outcome-selected replacements. Actual completion is reported in the run ledger.
Run
Time (s)
Req.
Applied
L/I/E/C
Usage ($)
Bound ($)
Status
E1+
2.0
20
20
0/0/0/0
0.0000
0.0000
complete
E2-
2.0
20
20
0/0/0/0
0.0000
0.0000
complete
H01-
90.0
900
900
0/0/0/0
0.0000
0.0000
complete
H01+
90.0
900
900
0/0/0/0
0.0000
0.0000
complete
H02-
90.0
900
900
0/0/0/0
0.0000
0.0000
complete
H02+
90.0
900
900
0/0/0/0
0.0000
0.0000
complete
Appendix
Table 6: All attempted runs. L/I/E/C denotes late, invalid, error, and cancelled requests. Usage and per-run conservative bounds are derived from the paid-request ledger; free/engineering runs make no paid calls. These are not per-run invoices. Times are observed simulation horizons.
Run
Wolf AUC
Founder surv.
R births
R deaths
R arrivals
Signals
P01-
3.3617
0.400
141
103
8
0
P01+
3.1830
0.500
151
106
8
191
P02-
3.8189
0.386
149
121
8
0
P02+
3.3332
0.457
143
104
8
169
P03-
3.1411
0.429
147
104
8
0
P03+
3.0703
0.457
135
104
8
222
Appendix
Table 7: Secondary descriptive outcomes for all 12 main attempts. Wolf AUC is normalized by initial wolf abundance; rabbit founder survival is the fraction of initial rabbit IDs remaining at the fixed horizon. No secondary causal mechanism is identified and no extra hypothesis test is performed.
Pair
On AUC
Off AUC
On − off
pair-01
1.6974
1.7245
-0.0271
pair-02
1.6762
1.6677
0.0085
pair-03
1.6028
1.5810
0.0218
pair-04
1.6469
1.6192
0.0277
pair-05
1.6734
1.6774
-0.0039
pair-06
1.7273
1.7530
-0.0258
Appendix
Table 8: Handwritten reference outcomes for 6 seed blocks. These are descriptive normalized rabbit-AUC values with immediate, zero-provider-latency decisions. The policy ignores received signals; its on/off differences can reflect emission cost. These runs are excluded from the main estimate and are not a matched comparison of model capability.
Figure 5: All 6 complete main-study seed pairs (12 runs), grouped by the frozen habitat seed. Each panel compares the two recorded arms for one initial state; provider responses and subsequent stochastic paths may differ. These descriptive traces do not establish ecological cycles or message-use mechanisms.
Mechanism
Main-study setting
Rabbit recruitment
Maturity 8 s; minimum energy and water 30 each; global crowding scale 160 rabbits; hard birth cap 250; local crowding disabled.
Rabbit birth cost
Parent loses 10+12f energy; birth or crowding rejection resets cooldown to 16(1.3−0.6f) s. Newborn energy/water: 60/80.
Metabolism 1.2 energy/s; a kill adds 40, capped at 100; movement energy charge is zero.
Movement
Rabbit speed 1.05+0.65gs−0.16gt tiles/s; wolf speed 1.6 tiles/s. Forest multiplies speed by 0.77; a wolf’s eating cooldown adds a factor 0.35.
Detection and capture
Rabbit sight radius 4+⌊6gv⌋ ; wolf sight radius 9, or 2 for a rabbit hiding in forest. Capture distance <0.8 ; eating cooldown 10 s; interference coefficient 1.
Appendix
Table 9: Effective parameters governing movement and demography. Trait symbols gs,gv,gt,f denote speed, vigilance, thrift, and fertility. Population caps constrain births, not immigration.
Run
Obs.
Any
%
Danger
Food
Follow
R req/min
W req/min
P01-
440
0
0.0
0
0
0
2.65
10.88
P01+
439
148
33.7
105
46
0
2.44
11.49
P02-
403
0
0.0
0
0
0
2.46
9.88
P02+
440
121
27.5
97
25
0
2.56
10.98
P03-
441
0
0.0
0
0
0
2.56
11.67
P03+
439
162
36.9
146
18
2
2.70
11.89
Appendix
Table 10: Signal exposure and population-normalized opportunities. +/− denotes signals on/off. Type columns count rabbit observations containing each kind and are nonexclusive; help is absent. The final columns are rabbit/wolf requests per living animal-minute, not per wall minute or per eligible adult.
Run/species
Obs.
Q (sim s)
Return n/N
W (wall ms)
Age n/applied
Age (sim s)
P01- R
440
18.24/19.49
438/440
112.5/172.5
438/438
0.102/0.153
P01- W
439
3.46/7.80
439/439
107.5/154.9
438/438
0.101/0.152
P01+ R
439
19.03/22.98
439/439
116.0/170.9
439/439
0.102/0.153
P01+ W
439
3.11/7.63
438/439
107.3/152.1
435/435
0.101/0.152
P02- R
403
17.36/24.82
398/403
112.9/161.8
398/398
0.102/0.153
P02- W
453
3.64/9.44
451/453
108.6/156.8
446/446
0.101/0.152
Appendix
Table 11: Per-run/species timing: R denotes rabbits and W wolves. Q is eligibility-to-dispatch queue delay; W in the time header is provider-adapter elapsed wall time. Time entries give median/95th percentile. Queue values are available for every listed observation. Return n/N explicitly excludes missing latency from quantiles; Age n/applied gives the matched-age denominator among applied requests. These are descriptive conditional distributions, not independent experimental samples.
Run
Horizon (sim s)
On (elapsed s)
Off (elapsed s)
P01
90.0
90.067
90.106
P02
90.0
90.110
90.064
P03
90.0
90.107
90.056
P04
90.0
90.091
90.106
P05
90.0
90.183
90.122
P06
90.0
90.051
90.062
Appendix
Table 12: Recorded elapsed seconds for every original attempt. P: paid main; H: offline handwritten control; E: engineering check. On/off denote the signal mechanism. The horizon column is simulation time; the other columns are real UTC intervals. A dash means no such arm was attempted under that label.
Agent-based modeling simulates how interacting, heterogeneous entities and their environments produce patterns in social, economic, biological, and engineered systems.
Recent growth in compute power has made it increasingly feasible to use large-scale agent-based models to simulate complex adaptive systems. A central difficulty is that such models contain many local rules and parameters, where small changes can lead to runaway behaviour, population collapse, or saturation at artificial bounds. We study this problem in a continuous predator-prey system where sheep and wolves are active agents with local sensing, internal energy, and recurrent neural network-based controllers. We ask whether environmental and demographic parameters can be tuned so that the resulting population dynamics resemble classical Lotka-Volterra cycles. We optimise these parameters with a feature-based loss that rewards sustained oscillations, phase lag, bounded populations, and long-term persistence, first for random controllers and then for evolved controllers in a more naturalistic setting. The model is implemented in ABMax, a JAX-based agent-based modelling framework that enables efficient batched simulation on hardware accelerators.
Corinna Mandl, Siddharth Chaturvedi, Marcel van Gerven
Department of Machine Learning and Neural Computing Donders Institute for Brain, Cognition and Behaviour, Radboud University Nijmegen, the Netherlands
Communication is typically understood as indication: signals that transfer information from sender to receiver. We present a minimal predator avoidance task in which pairs of evolved CTRNN agents use communication for robust survival, and in which agents hear their own vocalizations, as in natural systems. Across 112 perfect-fitness agents from over 2,000 evolutionary runs, three dominant strategies emerge (accounting for 81% of agents): safety calling (39%), where agents signal from safe cover; alarm indication (22%), where agents vocalize when a threat is present without relying on self-hearing; and self-regulatory calling (20%), where agents depend on hearing their own call to sustain escape behavior. Self-hearing dependency is common among agents that call during an active threat (47%), but rare among agents that call only after reaching safe cover (10%; p < 10^-4). This pattern is consistent with a difference in causal order: safety callers act then communicate, while self-regulatory callers communicate in order to act. Removing self-hearing selectively impairs self-regulatory callers (fitness 0.40) while safety callers remain functional (0.90; p < 10^-9). These results show that communication can evolve to serve the caller's own behavioral regulation, not just information transfer to others.
AI agents depend on repeated decisions to accomplish tasks, with large language models (LLMs) commonly serving as their core decision makers. The recent emergence of Jev offers an opportunity to make these decisions faster and at lower cost. However, it remains unclear whether such local savings can be translated into efficient and reliable task completion. In this paper, we investigate this efficiency transfer across five agents in two representative application settings, embodied tasks in ALFWorld and mobile app use in AndroidWorld. Across 4,500 task executions, we jointly evaluate task effectiveness, decision-level and task-level resource use, and execution behavior. Experimental results show that Jev integration substantially reduces decision cost and latency across all five agents, with task-level gains varying by workflow. Task success remains close to baseline levels on ALFWorld and declines more substantially in two mobile agents. Auxiliary reasoning and additional interactions absorb decision savings, while failed recovery and premature completion limit successful delivery. Mean cost per successful outcome increases in one agent even as spending per attempt falls. These findings highlight the importance of preserving progress, supporting recovery, and verifying completion when integrating efficient decision models into agent workflows. Their practical value should be assessed through successful task delivery, recovery behavior, and resources per delivered outcome.
Yuchen Ling, Shengcheng Yu, Zhenyu Chen +1
Nanjing University · Technical University of Munich