Plateau
Semantic circuit breaker for AI agents
A detector that trips on semantic information stagnation rather than textual repetition, so autonomous agent loops can be stopped when they stop making progress — not merely when they start repeating themselves.
The problem
An autonomous agent that has stopped making progress will often keep producing output that looks like progress. It rephrases its last conclusion, tries a variation of the tool call that already failed, restates the plan in new words. Tokens keep being spent; information stops accumulating.
The standard guard against this is repetition detection over the text. That guard fails in both directions:
- An agent that paraphrases itself is stuck, but the strings differ, so nothing trips.
- An agent legitimately iterating on a structured format produces similar text while genuinely progressing, so the breaker trips on healthy work.
The signal people actually want is not "is it repeating words" but "is it still learning anything."
The approach
Plateau measures semantic stagnation. Each step's output is embedded, and successive steps are compared in embedding space rather than as strings. When new steps stop moving the agent to a meaningfully different position, the breaker trips.
Two design choices carry the work:
Offline embeddings
Comparison runs on local sentence embeddings, not an API call. A circuit breaker that depends on a network round trip per step adds latency and a new failure mode to the exact loop it is supposed to protect.
A self-calibrating baseline
There is no universal threshold for "too similar." What counts as normal step-to-step movement differs by task, by model, and by prompt. So the threshold is not a constant — Plateau calibrates against the agent's own observed baseline and trips relative to that.
This is what makes it usable without per-task tuning, and it is the part that took the most iteration.
Validation
Testing a detector whose ground truth is fuzzy is its own problem: there is no labelled corpus of "stuck agent loops." Validation is a 121-test suite, plus a live two-machine demo run at the hackathon — a distributed setup rather than a single-process happy path.
Shipped as the installable package plateau.
Status
Built for the VIT Chennai Frontier AI Hackathon, July 2026. Shipped.