Agent skill¶
The aster agent skill is the agent-facing form of the bench. It is deliberately thin: it teaches
an agent when to use the bench, where the APIs are, and which evidence rules must survive the loop.
What is in the skill¶
skills/aster/
├── SKILL.md # trigger conditions, loop, non-negotiables
├── references/
│ ├── api-map.md # task → API index
│ ├── aedt-contract.md # guard/AEDT configuration + run lifecycle
│ └── evidence-rules.md # labels, kept artifacts, re-scoring
└── scripts/
└── dry_demo.py # runnable starter, no AEDT
Install¶
With the library installed:
Without installing, copy the folder into the agent's skill directory:
For other agent platforms, keep the same layout (SKILL.md + references/ + scripts/) and point
the agent at the folder. The YAML front matter carries the skill name and description; the body
is the instruction.
Try it¶
The script exercises the full dry loop: parameter space → Latin hypercube → evaluations → Pareto archive.
What the skill will not do¶
- It will not present a
PREDICTEDscore as measured. - It will not drop a failed run silently.
- It will not drive the web GUI (not published yet).