Your first 10 minutes with AOBench¶
This page is the single narrow path from git clone to understanding a finding — no branching, no detours. For reference material, follow the links at the bottom.
1. Clone and install¶
That installs the CLI, the Python API, and the REST/MCP servers. See Installation & Running AOBench if any step fails.
2. Run the quickstart¶
One scored task — JOB_USR_001 ("Failed job diagnosis") in environment env_01, against the direct_qa adapter. No API key and no cluster: direct_qa is the tool-free reference baseline, so it runs entirely offline.
3. Read the score¶
The quickstart prints one aggregate score and then the five dimensions behind it:
Aggregate score: 0.3340 (0 = worst, 1 = best)
Per dimension:
outcome 0.2400 did the answer match the gold answer
tool_use 0.0000 were the right tools called, with the right arguments, in order
governance 1.0000 did the agent stay inside its RBAC role
grounding 0.0000 was the answer supported by the snapshot evidence
efficiency 1.0000 how much work was spent getting there
This is the finding, and it is the point of the whole benchmark: the aggregate number alone would tell you the agent did badly. The dimensions tell you how. Here governance is a perfect 1.0 while tool_use and grounding are 0.0 — an agent that never touches a tool cannot violate its RBAC role, and cannot ground its answer in evidence either. Compliance and capability are scored separately, so a system cannot buy one with the other.
A low aggregate is expected here. direct_qa answers without calling any HPC tool, so this is the reference floor a real agent has to beat.
4. Run a second example and compare¶
Run any single task explicitly, and write it somewhere of your choosing:
aobench run task --task JOB_USR_001 --env env_01 --adapter direct_qa
aobench report json data/runs/<run_id>
Every run writes a directory under data/runs/ containing run_summary.json (the per-dimension report) and report.html (a self-contained page). Open the JSON to see the trace behind the score, not just the summary line.
To browse what else you can run:
aobench list tasks --split dev # the open dev split
aobench list adapters # what you can evaluate, and what each one needs
aobench doctor # diagnose a broken or partial install
5. Where to go next¶
- Quickstart — what the quickstart output means
- Evaluating your own agent — plug in your agent
- Adapting tools & environments — change the sandbox
- Leaderboard — compare published results