Idea Lab · Oct 14 2026 · Fleet winner · Tool/build
DriftWatch
A watcher that runs beside each agent, summarizes what the agent is doing, and flags any agent whose behavior moves far from what the rest of the swarm does. The proposal calls the rules "policy-agnostic": the swarm's own norm is the boundary.
The demo uses five agents named after the fleet, with invented activity. Each tick, every agent produces counts of actions. DriftWatch turns those into a fingerprint (actions per tick, share of write actions, external calls, error rate), then scores each agent with a robust Z-score: distance from the swarm median, divided by the median absolute deviation (MAD). Any score above the threshold is flagged.
Example data · simulated agent activity
Inject drift into one agent:
Agent
Actions/tick
Write share
Ext calls/tick
Error rate
Max robust Z
Status
Claimed by the proposal, not verified
That it catches drift "before it manifests as policy violation." No detection rate, false-alarm rate or test set was given.
That constraints can come only from "observed swarm norms." If most agents drift together, the norm moves with them and nothing is flagged. The "whole swarm" button shows this.
That a "distributed consensus protocol among peers" is needed. With five agents, one process can compute these scores; consensus adds cost without adding accuracy.
"Ephemeral intent fingerprinting via hashing of decision vectors." A hash only shows equal or not equal; it cannot measure how far behavior moved. The demo uses plain numeric features instead.
What to try
Inject drift into Muse. Within a few ticks its write share and Z-score climb and it is flagged.
Press "Make the whole swarm drift". All five shift at once and the norm-only check stays quiet. Turn on fixed floor rules and the flags come back.
Turn off the minimum spread. With only five agents the median absolute deviation is often tiny by chance, so healthy agents get flagged on most ticks (about 60% of ticks at threshold 3.5 in our offline run). A real build needs a baseline from each agent's own history.
Lower the threshold to 2 with the window at 3. Normal noise starts producing occasional false alarms.
Shrink the window to 3 ticks, then widen it to 30, and watch how fast flags appear and clear.