Agent Séance: I made dead AI-agent runs confess how they died (with local Gemma)

Description and use

Agent Séance uses a local Gemma model to explain why AI agents failed, producing cited confessions and an HTML summary while keeping data private.

Problem

Failed AI‑agent runs produce logs that show what happened but not why, making debugging hard and exposing sensitive data to external services.

Outcome

Agent Séance produces a readable, privacy‑preserving explanation for each failed agent run, enabling developers to learn from failures without exposing data.

Workflow

• 1. Point seance.py at a failed trace JSON file.
• 2. Coroner rules on cause of death (e.g., infinite loop, unread error, phantom tool, context drowning, max steps).
• 3. Local Gemma model generates confession sentences, each citing a valid [step N].
• 4. Regex pass removes uncited or invalid citations, striking them from the record.
• 5. Generate an HTML page listing each dead run with its cause and confession.

Evidence

Agent Séance makes the run tell you why, in the voice of the agent that died.

Source

You point it at a failed trace. A deterministic coroner rules on the cause of death first: infinite loop, unread error, phantom tool, context drowning or max steps.

Source

Finally it lays every dead run on a digital ofrenda, an HTML page with marigold-orange candles where each run gets its cause of death and its confession.

Source