| # Agent Reliability Eval |
|
|
| A short, runnable notebook on evaluating agent reliability along two axes |
| scored separately: **tool-call accuracy** and **hallucination rate** |
| (groundedness of the final answer against what the agent's tools actually |
| returned). |
|
|
| Open `agent_reliability_eval.ipynb` — it runs end to end with no API key |
| and no external dependencies beyond `nbformat`/`nbclient` if you want to |
| re-execute it; the shipped copy already has outputs baked in. |
|
|
| ## Why two metrics instead of one |
|
|
| End-to-end accuracy hides two different failure modes behind a single |
| number: an agent that calls the wrong tool but gets lucky, and an agent |
| that calls the right tool but pads its answer with unsupported specifics. |
| This notebook scores three mock agent policies with the same harness to |
| show it actually discriminates between those failure profiles, not just |
| pass/fail. |
|
|
| It also documents, on purpose, a real blind spot in the grounding check |
| used here — a fabricated qualitative claim the simple heuristic misses — |
| because pointing at your own eval's limitations is more useful than |
| hiding them. |
|
|
| ## Companion project |
|
|
| [`auditable-agent-demo`](https://github.com/kernalseed/auditable-agent-demo) |
| applies the same "don't trust output without checking how it got there" |
| principle to a spec-bound, auditable compliance-triage agent with a full |
| tool-call and rule-citation trail — the production-shaped version of the |
| idea this notebook explores in miniature. |
|
|
| ## Files |
|
|
| - `agent_reliability_eval.ipynb` — the notebook (run it, or read the baked-in outputs). |
| - `reliability_eval.py` — the underlying logic (tools, mock policies, eval functions), importable on its own. |
|
|
