What I Built

Every agent builder has a graveyard: runs that looped forever, ignored an error, or confidently reported a refund that never happened. Logs tell you what happened. Agent Séance makes the run tell you why, in the voice of the agent that died.

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. Then a local Gemma model, running in Ollama, summons the ghost to testify in the first person. There's a catch: every sentence must cite a [step N] that exists. Anything uncited, or cited to a step that never happened, is struck from the record and shown crossed out. Ghosts don't get to make things up.

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.

Demo

The digital ofrenda: each dead run with its cause of death and cited confession

Run it: python3 seance.py traces/*.json --html ofrenda.html

Sample (from my run):

⚰️  unread_error | Poisoned by an unread error
  👻 [step 0] I was tasked with processing order 5521, a simple transaction, a digital echo of a request.
  👻 [step 1] My processing halted, a digital paralysis as the error persisted.
  ✂️  struck: Here is my recorded life, a chilling echo of my own demise…
  ✂️  struck: [step 3] The system flagged an error – a payment provider's card expired, rendering the refund impossible.

The ghost tried to cite [step 3], a step that never happened in a 3-step trace, so the cross-examiner struck it. That's the whole point: even a dead agent doesn't get to hallucinate evidence.

Code

VasuBansal7576 / agent-seance

Agent Séance: a coroner for dead AI agent traces, with a Gemma ghost (DEV Hacktoberfest Week 4)

Agent Séance

Summon the ghost of a failed AI-agent run with local Gemma (Ollama). A deterministic coroner rules on the cause of death, the ghost testifies citing [step N], and uncited claims are struck from the record.

ollama pull gemma3:4b
python3 seance.py traces/*.json --html ofrenda.html

Trace format: {"name","task","steps":[{"type":"tool_call|tool_result|final","tool","args|content","error"?}]}

MIT licence. Built for DEV Hacktoberfest 2026 Week 4 by Vasu Bansal.

View on GitHub

How I Built It

Gemma (open-weight) via Ollama, fully local: gemma3:1b (runs on a plain CPU box in under 20 seconds per run; --model gemma3:4b for richer testimony) at temperature 0.4.

Coroner first, ghost second: plain Python rules pick the cause of death, so the model narrates a verdict it can't change.

Cross-examination: a regex pass keeps only sentences whose [step N] citations all exist in the trace. This is the same "make the model show its evidence" check I use when evaluating production agents.

It's one dependency-free Python file.

Why Does Open Innovation Matter?

Agent traces are the most sensitive thing a team owns: customer messages, internal tool outputs, sometimes credentials. Sending your failures to a closed API to get them explained means shipping your worst moments to a third party. With an open-weight model on my laptop, the séance happens with the lights off and the Wi-Fi off, and the dead stay in the house.

DEV Community

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