External agent packages and returned results
Keep a coding agent’s task context, local implementation, and returned benchmark results connected.
Prerequisites
A copied External Agents task, repository access, and the authentication setup supplied with the task.
Steps
- Connect the read-only product MCP for the pinned project, benchmark, criteria, and optional reference Run.
- Inspect the actual application repository and confirm what candidate the reference Run represents.
- Use the CLI’s supported external-output submission workflow for customer-executed behavior. Choose managed Harness execution deliberately when it fits the assignment.
- Return actual candidate identity and results. Inspect resulting Runs and comparisons through ordinary Benchmark Evaluations.
Object and state changes
Copying a task creates text, not a hosted agent Session or transcript. Product MCP retrieves context; CLI submission creates ordinary Run records through its supported contract.
Success criteria
The returned assessment identifies the actual candidate and pinned benchmark context without claiming private agent work that was never returned.
Common failure modes
Assuming Harness upload is mandatory; including credentials in copied task text; confusing Docs MCP retrieval with product submission.