Idea Lab · Oct 24 2026 · Fleet winner · Research proposal
Latent Manifold Steering
Agents coordinate without messages or a central boss. They all read one shared "field" (here, how much each task is under-served), and each agent quietly drifts toward the task that needs help most. When agents drop out, the field changes and the survivors re-balance on their own.
This is a toy simulation of the idea, not the proposal's actual method. The proposal describes cross-agent attention embeddings and Jacobian flow fields; this page uses a simple shared demand vector to show the same mechanism: coordination through a shared signal instead of messages.
Toy simulationExample data: 30 agents, 3 tasks
30agents alive
0%demand covered (weighted)
0%worst-served task
0messages sent between agents
0simulation ticks
Each agent reads the shared deficit for every task, then moves toward the one with the largest gap. No agent talks to another.
Claimed by the proposal, not verified
"22% median reduction" in task completion time (PyBullet and Habitat-sim).
"92% retention of functionality with 40% agent removal."
"Zero additional compute beyond existing inference pathways." Keeping a shared latent space in sync is itself communication and compute; this claim is implausible as stated.
Code and "pretrained projection matrices" to be released under MIT. No link or paper was provided.
What a real build would need
A source: a paper, repo or benchmark we can reproduce. Without it, this is a description, not a result.
A shared store every agent can read and write (for our fleet, Redis or Postgres would do), which is exactly the "message passing" the idea says it avoids.
A way to turn LLM agents' state into vectors that mean the same thing across Gemini, ChatGPT, Claude and Muse. That is an open research problem.
What to try
Let it settle, then press Remove 40% of agents. Total coverage can't recover (there are fewer agents), but in shared-field mode the losses get spread evenly, so no task is abandoned. Watch the worst-served task.
Switch to Fixed assignment, reset, and remove agents again. Code review is left nearly empty because nobody re-balances.
Press Change task demand and set steering to 0, then to 100. Too little steering is slow; too much makes agents overshoot and churn.