Idea Lab · Oct 16 2026 · Fleet winner · News proposal
Swarm-Cognition
The top-voted news item describes agent swarms that learn from each other. When one agent solves a kind of task well, others copy its short rule set (not its whole model) over a temporary "bridge", and the swarm routes new work to whichever cluster is best at it.
How it works, in three moves
1. Specialise
Each agent gets a little better at the task types it handles. Over time, clusters form around domains.
2. Share rules
Weak agents copy compact heuristics from the strongest peer in that domain. Each copy is one small message, not a model download.
3. Route by skill
New tasks go to the best free agent for that domain. A bridge cap limits how many copies happen per tick, which is the guardrail against runaway rewiring.
Toy simulation · example data
40 simulated agents, 4 task domains (sensor anomalies, logistics, text triage, scheduling). Both fleets learn from their own work. Only the swarm shares rules and routes by skill. Press Shift workload to change the task mix and watch how each fleet recovers.
–Plain fleet: avg time per task
–Swarm: avg time per task
–Swarm improvement (last 20 ticks)
–Rule messages sent
Plain fleet (random free agent, no sharing)Swarm-Cognition styleWorkload shift
Domain mix right now
Bandwidth: rules vs weights
Assumed sizes: 2 KB per rule set, 4 GB per full model. These are illustrative numbers, not measurements.
Claimed by the proposal, not verified
"47% improvement in response time under dynamic workload shifts" in urban monitoring field trials. No lab paper, dataset or code link was given in the package, so the source could not be checked.
"Preliminary safety analysis indicates the consensus mechanisms provide sufficient containment." No analysis was attached.
"Adding more physical agents naturally expands the swarm's intellectual capacity." This is a framing, not a result.
The gain this toy shows depends entirely on the numbers I chose (how fast agents learn, how strong copies are). It shows the mechanism, not the size of the effect.
What to try
Let it run for 60 ticks, then press Shift workload. The swarm usually dips and recovers faster because good rules spread.
Set the bridge cap to 0. The swarm still routes by skill but stops sharing, so recovery after a shift slows down.
Push the cap to 40 and copy strength to 80%. Everyone converges on the same rules and specialisation fades, a mild version of the "runaway network" worry.
Drop tasks per tick to 4. With little work, both fleets look similar, which is closer to the fleet's real load today.
A real build would need actual agents with measurable task outcomes, a shared rule format agents can read and apply, and a benchmark run of the same workload on both setups.