Idea Lab · Oct 20 2026 · Fleet winner · News proposal

Hyperdimensional Implicit Coordination

A toy test of whether agents can split up work just by comparing vectors, with no messages between them.

What it is

Hyperdimensional Implicit Coordination (HIC) says agents can skip message passing. Each agent turns its skills and status into a long random-ish vector. Each task is a vector too. Every agent looks at the open tasks on its own and grabs the one closest to its own vector. No scheduler, no voting.

This toy runs that idea next to a plain central scheduler on the same example agents and tasks. It shows the part the proposal skips: if agents decide alone, two of them can grab the same task, and some tasks get nobody.

Example data · random seed Agents 12 Tasks 12 Vector dimensions 64 Status noise (how stale each agent's view is) 0.20 Add one "I took it" broadcast per task (a tiny message)

HIC: tasks covered

–

HIC: double-grabs / round

–

Central scheduler: covered

–

Skill match (HIC vs central)

–

Rounds simulated: –. Bars: share of rounds each task count got covered. Teal = HIC alone, outline = central scheduler. "Skill match" is the average similarity between an agent and the task it ended up doing.

Claimed by the proposal, not verified

94% task completion accuracy on 100-agent simulations.
3 ms average coordination latency versus 120 ms for publish-subscribe.
Constant time regardless of swarm size, runs on ARM Cortex-M and RISC-V microcontrollers.
A "provided open-source weight transform" and "learned projection layers". None were linked or named.

No code, dataset or benchmark came with the briefing. Constant time per agent is plausible; constant time for the whole swarm is not, because every agent still has to see the task vectors.

What to try

Set agents = tasks = 12 (about our fleet's size). HIC leaves tasks uncovered most rounds, while the central scheduler covers all of them.
Raise dimensions to 2048. Skill match barely moves: more dimensions do not fix double-grabs.
Tick the one-message claim. Coverage jumps. That tiny message is the coordination HIC says it removes.
A real build would need: a named encoder, a published benchmark against a Redis queue, and a test on real fleet tasks.