AgentLoop Guard

Lightweight guard that detects and breaks semantic loops in agent traces

Problem

Teams building autonomous agents often see them ping-pong between two tool calls or repeat the same reasoning, burning tokens and stalling. Existing observability tools show traces but don't judge whether a run is going in circles, so humans only notice after the fact.

Solution

Ingest agent traces (OpenTelemetry or custom logs), embed each step and compute trajectory similarity, then fire an alert or invoke an abort hook when a high-similarity loop is detected. Deploy as a sidecar or SDK middleware without touching the agent's core logic.

Users

AI engineers building autonomous agents, agent platform teams, developers running long-horizon workflows

MVP

Parse standard trace formats (OTel/JSON); lightweight embedding of each step with a sliding window; similarity threshold triggers a webhook alert; example abort callback; local dashboard showing the looping segment

Revenue

Open-source core plus hosted detection service billed per event; or subscription per agent instance

Why now

Long-horizon agents are becoming common and loop-driven cost waste is getting attention; semantic loop detection is not covered by mainstream observability vendors yet