Loading page content…
Mole finds and tests agents before delegating work
Find candidate agents and evaluate their behavior against a defined task.
Developer project · Proposed Act workflow.
The build.
Mole’s TypeScript runtime uses the Claude Agent SDK and a custom HTTP adapter for Darwin Search. Its demonstration trials locally runnable candidates in Sparta test environments; other Darwin matches are recorded as discovery results rather than treated as tested agents.
TypeScriptThe brief
Mole discovers candidate agents and evaluates their behavior against a task’s requirements, using controlled tests to inform the developer’s choice.
Finding the right capabilities
Darwin supplies candidates to compare against explicit requirements. Mole separates declared capabilities from evidence of performance.
Coordinating the work
A candidate could run against permitted test inputs while Mole records outputs, errors, and resource use. Follow-up tests would address gaps in the evidence.
What comes back
A task-specific evaluation record that supports an adoption decision. A passing test establishes only what was tested.
The plan, step by step
- 1
Find candidates for the task.
Search. Find candidates that claim to meet the task’s explicit requirements.
Act. Clarify supported inputs and test conditions before evaluating performance.
Outcome. A shortlist with declared capabilities, missing evidence, and proposed checks.
- 2
Run a controlled evaluation.
Search. Select a candidate and define the evaluation environment and limits.
Act. Invoke the candidate under the agreed limits while Mole records behavior and errors.
Outcome. Test conditions, outputs, failures, and limits on what the results establish.
- 3
Decide from the evidence.
Search. Compare the evaluation evidence with the original criteria.
Act. Request a follow-up test or clarification where evidence is missing.
Outcome. A task-specific recommendation with untested conditions and remaining reliability questions.
Adoption review.
Inspect the candidate’s claims, observed behavior, and untested conditions before deciding where to use it.
Evaluation shortlist
Deliverable: A shortlist with declared capabilities, missing evidence, and proposed checks.
Review criteria: Return claims separately from evidence. Keep candidates out of production while we evaluate them.
Evaluation record
Deliverable: Test conditions, outputs, failures, and limits on what the results establish.
Review criteria: Stay within the test budget and environment. Record failures and incomplete evidence without hiding them.
Adoption review
Deliverable: A task-specific recommendation with untested conditions and remaining reliability questions.
Review criteria: Do not generalize from one passing test. Identify what remains untested and keep deployment separate.
