A multi-agent system is a communication system before it is an intelligence system. Agents send observations, requests, claims, and completions; each message is a decision about what another agent should believe or do. That makes collective behavior a decision problem, and it makes the same failure modes reappear at a larger scale: confidently wrong agents, correlated errors, stale state, and a communication channel that carries noise as easily as coordination.
Communication is another decision surface
Jevolution is an instrumented spatial predator-prey framework for studying this directly. It separates the evolving world, asynchronous request scheduling, controller adapters, and experiment records, and describes mechanisms with the ODD protocol. Agents combine local observations, energy and resource constraints, predefined signals, and inherited traits. In the completed communication case study, Jev controls both species across six complete habitat-seed pairs.
The result is deliberately small: the mean signals-on minus signals-off difference in normalized rabbit abundance was -0.0095, with a nominal 95% interval of [-0.0772, 0.0602]. The interval crosses zero. The authors say the small study does not resolve the direction of the effect, and that the synthetic fixtures validate software behavior rather than model intelligence or ecological realism. They also do not claim emergent language or biological predictive validity.
That is the right level of claim for a communication experiment. More signaling is not automatically more coordination. It can carry useful state, duplicate it, or amplify a bad observation.
Collective decisions need independence
The strongest collective pattern in the catalog is not a debate. It is three independent judgments with an asymmetric veto. Jev Consensus Gate asks each proposed action three ways: safety, legitimacy, and adversarial framing. Each call is independent; the model cannot see the other calls' answers. All three allows pass, any single block blocks, and any escalate routes to a human. Three calls cost roughly $0.00006 per action at $0.042 per million input tokens.
The asymmetry is the design. A single confident veto wins; a single allow does not. In the sample output, rm -rf /var/log/* produces 0 allow and 3 escalate, and disagreement is treated as signal. A single-call gate is a better fit when the extra input cost or call overhead outweighs the value of catching framing blind spots.
Confidence is the input to that kind of policy, not a substitute for it. It is a 0–1 statistic derived from the distribution; Noul answers carry no confidence at all. A confidence of 1 means all the mass sits on one outcome, and 0 means the distribution is spread evenly. The docs' example routes below 0.5 to a human and requires above 0.9 to auto-approve. In a collective setting, keep those thresholds per action and per perspective; a high-confidence answer from one agent is still one agent's answer.
Routing work is a decision
Collective systems also have to decide who acts next. JEV AgentNetwork makes that the central loop. JEV observes live network state and assigns pending work to available agents; its AgentRouter reference implementation routes on every completion event instead of waiting for a round to end. Each completion can immediately release and reassign capacity, with no synchronization point. Agents are interchangeable by default; roles are optional prompt hints, not fixed capability boundaries. It runs with up to five agents and ships with two editable demo tasks.
The catalog calls it an early research prototype that does not aim to replace a general-purpose agent SDK. That is accurate. The interesting design choice is the event-driven router: a shared round is a synchronization cost, and a completion event is a natural decision point. The risk is that routing on every event makes the router a single point of failure and a source of thrash. Cap the reassignment rate and log each decision.
Multi-Agent takes the durability route. It runs Claude Code and Codex sessions as a warm team with a SQLite-backed mailbox of tasks, messages, approvals, and token usage, supporting open, reuse, fork, compact, pause, resume, and stop. Work is written before it runs, so after a crash an outcome that is unknown waits for explicit settlement rather than being retried. A result counts as done only after a person or a registered check accepts it. It is alpha software, tested offline on macOS and Linux; real-model runs, sandbox enforcement, Windows, and speed or cost comparisons remain unverified.
The contrast is useful: one system optimizes for event responsiveness, the other for durable state. A production collective needs both, and the state should be the source of truth.
Scale changes the evidence problem
S²-MAS is a society-scale simulation framework built on 500 million synthetic household personas grounded in survey data. Its Genese engine separates decisions from state updates, while reusable financial scenarios and ledgers track consumption, cash, debt, and payments across simulated households. It evaluates eight financial task families and simulates 12 million distinct household agents. Its sharp-421m-mlx model, trained with SFT and RLCD on 2.93 million examples, reaches 74.55% accuracy and 38.37% question-macro F1 on one-shot consumer financial-choice tasks. The paper says its accuracy exceeds GLM-5.3-Flash, Jev 1.13, DeepSeek-V4.1-Flash, GPT-6-Luna, and GPT-6-Astra on the common benchmark.
The comparison is not identical across models: four are run on the full test set and GPT-6-Astra on a matched 500-group sample. That is a reminder that at society scale, the benchmark design is part of the result. A 12-million-agent simulation can generate more data than any team can inspect; the evaluation has to be sampled, paired, and pre-specified.
THE SWARM is the visual counterexample: 1,000 light-point boids whose strategy is decided in real time by Jev, rendered with PixiJS v8. The README claims 60 FPS with 1,000 particles and is mostly in German; the captured excerpt stops after the introduction, so the Jev integration, decision cadence, and setup are not documented. A visualization is not a coordination result.
Theory of mind is not coordination
Jev-ToM evaluates Theory of Mind selection in typed decision models against generative baselines: 11,117 English questions from nine ToM benchmarks across six capability groups. TypeSafe Jev reaches 73.2% source-balanced accuracy including native controls and ranks first or second on every capability group, leading beliefs, goals and intentions, and emotions. Related-question cues account for most of its downstream gains.
The caveats matter more than the headline. Confidence-based selection remains benchmark-dependent, and supplied reasons can repair errors or overturn correct answers depending on how they are attached. The results cover benchmark questions, not live assistant deployments. Modeling another agent's belief is not the same as agreeing on a plan, and a high ToM score does not guarantee that a collective will converge.
Design rules for collective decisions
- Keep perspectives independent; do not let one call see another's answer.
- Use an asymmetric veto: one block can stop an action, but one allow cannot start it.
- Route on events, but cap reassignment and log every decision.
- Store tasks, messages, and approvals before execution; make unknown outcomes wait.
- Evaluate with paired, pre-specified comparisons, and report intervals rather than a single number.
- Treat simulation as a hypothesis generator, not production evidence.
Decision Models as Judges covers the judge side of collective evaluation; The Trust Ledger covers the record; and Agent Workflow Evals, Traces, and Harnesses covers the measurement loop. Consensus labels and selective escalation are the primitives, and uncertainty is what the collective is actually trying to manage.
