AgentTrance — Debug your runaway AI agent

When your AI agent goes haywire, reconstruct exactly what happened

Problem

A user provisioned a fresh Hermes agent on a headless Debian server, asked it to print its environment and $PATH using Z.ai's GLM model, and it spiraled into a trance-like state of gibberish output. The user posted the log on Hacker News asking whether the output meant anything at all — showing there is no way to replay, snapshot, or trace what triggered an agent's breakdown.

Solution

A lightweight sidecar recorder that logs each tool call, prompt, token usage, state transition, and final output into structured local storage. When the agent misbehaves, it flags the 'trance' segment and generates a readable timeline that shows where the drift started and which inputs triggered it.

Users

Solo developers and ops engineers running self-hosted AI agents; small teams using local models for automation.

MVP

Support 2-3 mainstream agent frameworks; log prompts, tool calls, and outputs to local SQLite; heuristic rules to flag anomalies (repetition, syntax collapse, loops); generate a readable anomaly timeline; fully local, data never leaves the machine.

Revenue

Open-core model with paid desktop/team edition; monthly subscription for multi-agent management and cloud sync; enterprise webhook alerts for agent anomalies.

Why now

Self-hosted agents and local models are exploding, yet there is almost no observability tooling aimed at LLM state itself. Traditional APMs don't model agent cognition — a purpose-built 'black box' is a clear indie gap.