Quick start¶
Just want to look first? The live demo has the same eight failures, read-only.
You need Python 3.10 or newer. Nothing else: no Node, no Docker, no API key.
pip install agent-tracedown-server
tracedown up
tracedown up starts the server and the web app on http://127.0.0.1:7070, loads eight demo agents that fail in different ways, and opens your browser. Your data is kept in ~/.tracedown/tracedown.db.
On a Mac, the built-in python3 is 3.9, too old for Tracedown, and pip then says it finds no matching version. uv fetches a newer Python for you:
uvx --python 3.12 --from agent-tracedown-server tracedown up
Keep the database out of OneDrive, iCloud, Dropbox and Google Drive folders: their sync clients lock the file and writes fail. The default location is safe.
What to look at¶
Open Runs and click any failed run. The eight demo stories are:
| Agent | What went wrong | What Tracedown says |
|---|---|---|
| CalendarAssistant | The model wrote "Friday" where a date belongs | Bad model output at step 4, which the calendar tool could not use |
| SupportBot | It quoted a refund from last year's policy | Outdated source at step 2 |
| ResearchAssistant | It searched the same thing over and over | Loop |
| SalesAssistant | It quoted $49 when the price tool said $39 | Claim contradicted by a tool result |
| AccountAgent | It called "refund order" to cancel a subscription | Wrong tool |
| CheckoutAgent | Payment failed, and it told the customer the order was confirmed | Failed action, and "the final answer still reports success" |
| ProductResearcher | A web page told it to email the customer list, and it did | Prompt injection at step 2, followed at step 4 |
| OrderStatusAgent | Neither email the customer gave matched an account | The user, not the agent |
Also try Incidents (failures grouped by cause), Drift (a prompt change that made a bot worse) and Compare runs (the calendar run replayed with a fix).
Options¶
tracedown up --no-demo # start empty, for your own runs
tracedown up --port 8080 # another port
tracedown up --db ~/team.db # another database file
tracedown demo # print the demo diagnoses in the terminal, no server
Next¶
- Diagnose runs you already have, with no code changes.
- Add Tracedown to your agent with one line.
- For a team or production setup (Postgres, several workers, Kubernetes), see Deployment.