CL-05 / Agent handoffs / proposed
An agent handoff you can inspect
What must a handoff contain so another agent can continue without inventing authority?
01 / One week
Make the smallest useful proof.
Define a handoff schema: task, evidence, permissions, unknowns and stop condition. Compare it against free text on 20 synthetic tasks.
02 / Four weeks
Stretch the experiment.
Add validation, replay traces and a human approval checkpoint for external actions.
03 / Materials
Start with what is available.
- $0–60; optional capped model API usage
- Skills to practise: JSON, evaluation design, prompt engineering
Illustrative materials only. Equipment purchases, expenses and access need human review.
04 / Privacy + boundaries
Choose a safe test fixture.
Use synthetic tasks and redact secrets. No autonomous sending, purchases or permission escalation.
05 / Success
Measure the change.
Every blocked action stays blocked; at least 18 of 20 handoffs preserve the stated next step
06 / Deliverable + review
Leave a usable handoff.
Schema, validation tests and scored evaluation set
Human review: Mentor reviews rubric and API cap. Humans approve consequential actions; agents can draft and test within the sandbox.
Before work enters the world.
- Pay, hours, eligibility, supervisor, location, agreements and legal review must be resolved before recruitment or work begins. Budgets are illustrative materials estimates in USD, exclude labor, tax and shipping, and are not spending authorization.
- Use synthetic or licensed public fixtures first. A human supervisor must approve any participant involvement, consent, collection and data lifecycle.
- Agents may draft, code, propose tests and document evidence. They cannot supervise interns, approve their own work or infer authority to send messages, buy materials or publish.
- PointCast publication is the programme’s intended context. Physical venues, partners and installations need separate confirmation; no external venue partnership is claimed.
- Credit every contributor with permission. Publish measured results and limitations only after human review; preserve a record of corrections.