Lab assignment
Coordinator and specialist agents
Coordinate three bounded specialists without losing provenance or hiding a coverage gap.
The brief
Your assignment
Create a small source set from public campus-sustainability pages or fictional excerpts, then prepare a short briefing. Use three specialists: source discovery, claim checking, and synthesis support. Run only independent discovery work concurrently; synthesis waits for the evidence it needs.
For the architecture path, write the three contracts and walk two synthetic traces by hand: one complete and one with a failed specialist. A build path may use Claude Code agents, but it must not upload private campus or student data.
Before you build
Set up a safe workspace
- Review the mapped Domain 1, 2, and 5 tasks.
- Create a small fictional or public source set.
- Define the fields for a finding, uncertainty, and source before assigning work.
Build in checkpoints
One observable behavior at a time
- 01Draw the dependency graph and tool-access matrix.
- 02Write a bounded contract for each specialist.
- 03Mark which discovery work may run concurrently.
- 04Walk a successful trace and then fail one specialist.
- 05Verify that synthesis retains provenance and names the coverage gap.
What you'll practice
- Give every agent one objective and only the tools it needs.
- Pass context and uncertainty explicitly.
- Retain claim-to-agent-to-source mappings through synthesis.
- Disclose partial coverage rather than presenting a failed team as complete.
Your lab files
Four files for each lab
- 01
decision.mdState the requirement, chosen pattern, rejected alternatives, and the reason for the tradeoff.
- 02
implementation/Keep the smallest working example, configuration, diagram, or synthetic-trace analysis.
- 03
evidence.mdRecord test cases, observed results, failure behavior, limitations, and evidence locations.
- 04
review.mdExplain what must be deterministic, what may remain model-directed, and where a human is required.
What to save
- A dependency graph and tool-access matrix.
- Three agent contracts with structured return fields.
- A claim ledger.
- One successful trace and one failed-agent trace.
Test bench
What good work includes
- An out-of-role tool attempt is denied.
- Trace evidence demonstrates genuine concurrency for independent work.
- Synthesis waits for required discovery.
- Every final claim maps to an originating agent and source.
- A failed specialist produces a named coverage gap.
Review questions
- When would one agent be better?
- How narrow is too narrow?
- Which recovery belongs to a specialist, and which belongs to the coordinator?
Strong solutions tend to
- Hub-and-spoke control with minimal tool sets.
- Structured uncertainty and claim-level provenance.
- Coordinator-owned decisions about cross-agent failure.
Partial example
What useful evidence can look like
This is one small example, not a complete solution or a template you must copy. Use the notes to see why the evidence is useful.
Open annotated example from claim-ledger.csv / one row
claim_id,agent,source,location,uncertainty,status
C-04,claim-checker,campus-plan.pdf,p. 12,"date may be outdated",verified- agent + source
- Keeps responsibility and evidence attached to the claim after handoffs.
- uncertainty
- Preserves a limitation instead of letting synthesis make the claim sound stronger.
Review your work
Use evidence from your work—not confidence alone. For each area, run the check and choose the strongest stage your work fully demonstrates.
Specialist roles and tools
Each specialist has one clear job, only the tools needed for that job, and a return contract the coordinator can use.
Compare the three contracts with the tool-access matrix, then attempt one tool call that should be outside each specialist’s role.
Needs another pass
Roles overlap, tool access is broad, or a specialist can return free-form results without the fields the coordinator needs.
Partly demonstrated
The roles differ, but at least one objective, tool boundary, or return field is too broad to test clearly.
Demonstrated
Every specialist has one bounded objective, a minimal tool set, and required result, uncertainty, failure, and source fields. Out-of-role calls are denied.
Dependencies and concurrency
Only independent work runs together, and the coordinator supplies earlier findings to work that depends on them.
Follow the dependency graph through the successful trace. Mark which tasks overlap in time and where a later prompt receives an earlier result.
Needs another pass
Dependent specialists run in parallel, or the design assumes that one specialist can automatically see another specialist’s context.
Partly demonstrated
The graph labels dependencies, but the trace does not prove that dependent work waited or received the required result.
Demonstrated
Independent discovery runs concurrently. Dependent work waits, and the coordinator explicitly passes the evidence it needs in a later invocation.
Claims and sources
A reader can trace every final claim through the responsible specialist to the exact supporting source.
Choose any final claim and follow its claim-ledger entry back to the originating agent, source, location, and stated uncertainty.
Needs another pass
The briefing has a general source list, or a handoff drops the source location or uncertainty attached to a claim.
Partly demonstrated
Some claims have complete mappings, but at least one final claim or handoff cannot be traced to exact evidence.
Demonstrated
Every final claim keeps its agent, source, precise location, and uncertainty through all handoffs and into the briefing.
Failure and coverage
A failed specialist changes the coordinator’s decision and the final report states what is no longer covered.
Run the failed-specialist trace and compare it with the complete trace. Identify the partial evidence kept, the recovery choice, and the named coverage gap.
Needs another pass
The coordinator presents complete coverage after a failure, or stops and discards useful results from specialists that succeeded.
Partly demonstrated
The failure is visible, but the coordinator has no clear retry, reroute, or disclosure decision and the missing coverage is vague.
Demonstrated
The failure returns useful context, successful findings remain available, and the coordinator retries, reroutes, or proceeds with a specific disclosed coverage gap.