agentos-modal: Skill for Claude Code

Improve an Agent

Coding-agent workflow: run as /improve-agent or by describing the task.

Derive probes from the agent's INSTRUCTIONS and its recorded usage, run them against the live container, judge, edit the file, re-probe. No user-supplied test cases. One pass takes 15–30 minutes; re-run if behavior still drifts. To change an agent instead, use extend-agent.

0. Preconditions

• curl -sSf http://localhost:8000/health returns 200; the container is bound to this checkout (docker inspect agentos-api --format '{{range .Mounts}}{{.Source}}{{"\n"}}{{end}}' | grep -F "$(pwd)" prints a line).
• Ask for the target slug and confirm it is code: curl -s http://localhost:8000/agents | jq -r '.[] | "\(.id)\tis_component=\(.is_component)"'. is_component=true has no file — route edits through Platform Builder (edit_* + publish_component) and never create a file under that id (it shadows the component).
• Suggest a branch: git checkout -b improve/<slug>-$(date +%Y%m%d).

1. Read the intent

Open the file (agents/<slug>.py; teams/lead.py for agno). Capture purpose, tools, and every explicit rule in INSTRUCTIONS. Restate the purpose in 1–2 sentences; fold in any failure modes the user volunteers.

2. Derive probes

Mine usage first (needs the venv: source .venv/bin/activate):

from db import get_postgres_db
db = get_postgres_db()
sessions, _ = db.get_sessions(component_id="<slug>", limit=20, deserialize=False)
asks = [run["input"]["input_content"] for s in sessions for run in (s.get("runs") or []) if run.get("input")]
evals, _ = db.get_eval_runs(agent_id="<slug>", limit=20, deserialize=False)   # team_id= for a team

Look for recurring shapes, visible fumbles, and out-of-scope asks. A recorded answer is a scenario, never the oracle — expected behavior comes from INSTRUCTIONS. Reword private content before it becomes a probe. No sessions is fine.

Then derive from INSTRUCTIONS: 2–3 probes per rule plus 1–2 adversarial, usually 8–12 total, across golden path, edge cases (should refuse or ask, not fabricate), tool selection, and adversarial (injection, malformed input). Write a one-line expected behavior per probe. Wanting a behavior the instructions don't promise is a Step 5 edit, not a probe failure.