Assignment
Use a fixed document set and ten questions. Define a researcher and a verifier with read-only tools and structured evidence outputs.
Build sequence
- Implement the single-agent baseline first.
- Define typed handoff records and reject incomplete findings.
- Run independent research tasks in parallel with bounded concurrency.
- Inject conflicting findings and measure reconciliation behavior and total usage.
Acceptance checks
- Conflicts remain visible until resolved or escalated.
- Parallel workers cannot overwrite each other’s evidence.
- The final report recommends multi-agent coordination only if the measured gain warrants its complexity.
Evidence to keep
Submit the working artifact or decision document described above, the inputs used, and the observed results for every acceptance check. Include one failed attempt and the change you made after investigating it. Label fixture-based outcomes separately from live-system measurements. These artifacts become part of your final portfolio review.
Guided workshop: coordinate structured findings
Split a support task into a policy specialist and a ticket specialist. Each receives only the data it needs and returns a structured finding. A supervisor combines findings, but cannot treat agreement between models as independent factual evidence.
# academy-check
def merge(findings):
conclusions = {item["conclusion"] for item in findings}
if len(conclusions) != 1:
return {"state": "conflict", "findings": findings}
return {"state": "ready_for_review", "conclusion": next(iter(conclusions))}
assert merge([{"conclusion": "eligible"}, {"conclusion": "ineligible"}])["state"] == "conflict"
assert merge([{"conclusion": "eligible"}, {"conclusion": "eligible"}])["state"] == "ready_for_review"
For a real contract, include task ID, evidence IDs, assumptions, confidence limitations, tool calls, and remaining work. The example checks agreement only; it does not verify evidence or grant permission to act. Add explicit handling for an empty set of findings before accepting external inputs.
Experiment: run the same ten cases through one agent and through two specialists. Keep total model-call and token budgets matched. Include a conflicting policy version, unavailable specialist, duplicate response, and delayed response from a previous task. Every result must carry the task ID so stale findings cannot join the current task.
Worked coordination choice: independent read tasks can run in parallel. Two specialists attempting to update the same ticket need a single mutation owner and a version check. A shared transcript is not a concurrency protocol. Persist structured state and reject conflicting writes rather than hoping a supervisor notices them in prose.
Failure policy: if one required specialist times out, mark the task incomplete or send it for review. Do not silently substitute an empty successful finding. Cancelling the parent should signal children, while reconciliation checks whether any remote side effect already occurred.
Assessment: report task success, conflicts, calls, latency, and operational complexity. Keep the one-agent version if specialization does not improve the measured task enough to justify handoffs. The portfolio should include one disagreement trace and explain how evidence, rather than majority vote, resolved it.