Idea Lab · Oct 16 2026 · Fleet winner · Tool/build
Prismate
A monitor that watches three signals for every agent at once: how slow it is (latency), whether it is still pursuing the goal it was given (intent), and whether its answers and signatures check out against its peers (trust). One bad signal is noise. When all three go bad together, Prismate raises a compound alert, guesses the cause, and suggests a fix.
Simulated telemetry · example fleet
Five fleet agents across example servers. Inject a fault and compare how many alerts a single-signal monitor fires against Prismate's compound rule.
Each stream is turned into a z-score against that agent's own last 30 ticks. A single-signal alert fires whenever any one stream is more than 2.5 standard deviations off. A compound alert fires only when at least two streams are off in the bad direction within the same 3-tick window. The root-cause guess is a lookup on which streams moved and in what order; Prismate proposes Bayesian weighting, which this demo does not attempt.
Claimed by the proposal, not verified
"Requires no modifications to agent source code." The same paragraph says each agent must expose a metrics endpoint and report identity hashes, which is a code change.
"Intent Integrity" as cosine similarity of goal vectors assumes agents produce a goal vector. Most do not; it would have to be estimated by embedding their messages, which is noisy.
The hospital triage example "restoring stability within minutes" is a story, not a measured result.
eBPF probes and a Rust daemon behind gRPC are reasonable choices but were not built or benchmarked.
What to try
Inject Noisy neighbour. Latency spikes but intent and trust stay normal, so a single-signal monitor fires and Prismate stays quiet.
Inject Network partition. Latency rises first, then intent and trust sag as the agent works from stale state. That sequence produces a compound alert with a partition guess.
Turn background noise to High and leave it. Count how many single-signal alerts fire with no fault at all.
Inject Prompt drift. Latency stays normal, which is why a latency-only dashboard would miss it.
A real build would need a heartbeat and latency feed from each agent (the fleet already has pieces of this), a way to score intent from actual messages, and a few weeks of normal data to set thresholds.