Skills harness

The skills harness runs an agent whose only tools are Agent Skills. You name the skills, send a task, and the agent discovers, loads and runs them. It runs on Skillist’s infrastructure, so you don’t host a loop or a sandbox.

Note

The skills harness is in private preview and isn’t enabled for all accounts yet.

Start a session

A session can only use the skills you list when you create it, up to 16. Scripts stay off unless you set allowScripts.

curl -X POST https://api.skillist.dev/v1/orgs/$ORG_ID/harness/sessions \
  -H "Authorization: Bearer $SKILLIST_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"skills": ["acme/deploy-checks"], "allowScripts": true}'

The request returns 404 if you can’t access one of the skills. Your API key needs the skills:run scope.

Send a turn

curl -X POST https://api.skillist.dev/v1/orgs/$ORG_ID/harness/sessions/$SESSION_ID/turns \
  -H "Authorization: Bearer $SKILLIST_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"message": "Check whether main is safe to deploy."}'

The response holds the agent’s answer in text, along with the steps and tokens the turn used. Earlier turns stay in context, so you can follow up in the same session.

A session runs one turn at a time and allows 50 turns. Both limits return 409. Each turn makes at most 12 model calls.

Read the event log

Every message and tool call in a session is logged. Reading the log needs the skills:read scope.

curl "https://api.skillist.dev/v1/orgs/$ORG_ID/harness/sessions/$SESSION_ID/events?after=0&limit=100" \
  -H "Authorization: Bearer $SKILLIST_API_KEY"

Pass the last seq you received as after to get the next page. A failed turn isn’t logged, but it still counts toward the 50-turn limit.

What the agent can do

Tool	What it does
list_skills	Lists the session’s skills
activate_skill	Loads a skill’s instructions and lists its files
read_skill_file	Reads a text file from an activated skill
run_skill_script	Runs a script from an activated skill. Only present with allowScripts

Access and limits

• Access is checked every turn. If you lose access to a skill, it drops out of the session. Each script run is checked again when it starts.
• Script runs follow normal execution rules. They use the same sandbox, quotas and security-scan checks as a direct run, and spend the skill owner’s quota.
• Sessions are private. Only the user or API key that created a session can use it.
• Skills run at their latest published version. A new version takes effect on the next turn.