Agenttrace Harness Demo — Data skill for Claude Code

agenttrace-harness-demo

An example agent workflow, written separately from [AgentTrace](https://github.com/AjayMaan13/AgentTrace), that records its runs into AgentTrace through the **Connect your agents** steps.

Three Claude agents, built on the Claude Agent SDK, write an account health brief:

researcher  ->  analyst  ->  writer
(looks up the account's tickets and open deals)   (judges renewal risk)   (writes the brief and states the numbers)

A check written in code decides whether the brief is right: the open-ticket count and the open pipeline in the brief must match the data. The data is a small built-in company (`acct-1` with 40 tickets and 3 open deals, `acct-2` with 5 tickets), so nothing outside this repo is needed.

`--bug` plants a failure that shows up far from its cause: the researcher's ticket reader stops after the first 20 tickets and reports that page as if it were all of them. The writer then states the wrong count, the check fails, and AgentTrace can follow the number back to the researcher.

Run it

python3.11 -m venv .venv && .venv/bin/pip install -r requirements-dev.txt
cp .env.example .env        # then fill in .env yourself; it is gitignored
.venv/bin/pytest            # offline: no network, no tokens
.venv/bin/python -m account_health.run --offline        # stand-in agents: no Anthropic key, no tokens
.venv/bin/python -m account_health.run                  # a good run, on Claude
.venv/bin/python -m account_health.run --bug            # the planted bug: the check fails

Each Claude run makes three small calls (Haiku) and needs `ANTHROPIC_API_KEY`. `--offline` swaps the three agents for stand-ins that answer from the same lookups with a simple rule. It is for trying the AgentTrace connection first; the stand-ins' steps are labelled `offline-stand-in` so nobody mistakes them for Claude.

Connect it to AgentTrace

1. In AgentTrace, open Connect your agents and make a key that can send runs. Put it in .env