Agent Skills: Clous Agent Runs

Clous Agent Runs

Agent runs can be created through API responses or A2A.

Workflow

1. Decide API responses vs A2A based on caller identity.
2. Create the run with explicit input and any required object context.
3. Store run_id or call_id.
4. Poll until a terminal state or follow the stream URL.
5. Summarize outputs, artifacts, pending approvals, and failures.

Run Summary Contract

• Include run surface, id, terminal status, object context, artifacts, approvals, failures, and unresolved follow-up.
• If the request used dry-run, preview, or test mode, say which target materialization or mutation was skipped.
• Do not describe a run as completed until the terminal state is observed or the stream explicitly closes with a final result.
• Preserve trace ids, conversation ids, or tool call ids when available so the product team can replay or audit the run.

Run Inspection Contract

When inspecting an existing run, keep the evidence chain visible:

1. Record the run identifier, endpoint family, caller/workspace context, current status, and last updated time.
2. Separate final output, intermediate artifacts, tool calls, approvals, errors, and retries.
3. For failures, preserve the smallest useful raw error or status payload before interpreting it.
4. For pending approvals or human-in-the-loop waits, report the exact blocker and the next allowed action.
5. For completed runs, include what was verified and what still needs product, API, or user confirmation.

Guardrails

• Do not re-run or mutate a workflow while inspecting unless the user explicitly asks.
• Do not collapse A2A call_id and /v1/responses run_id; keep both names distinct.
• Do not expose employee, candidate, or customer data from artifacts when a redacted summary is enough.