▶ Demo: building your first agent team

Module 04 demo · ~4 minutes · prompt by prompt; run these yourself and you land in the same place

Repo structure & tests

client-project/

.claude/

agents/ (empty until today)

researcher.md (read and search only)

reviewer.md (read only)

research/

8 meeting notes, May through July

deliverables/

2026-08-24-intake-summary.md DRAFT

Boundary tests

detail not in report → session does not know ✓

asked again, more specifically → found ✓

asked to edit → cannot (no Write tool) ✓

.claude/agents/researcher.md (what Claude wrote)

---
name: researcher
description: Reads background material and
  produces short, source-cited summaries. Use
  for any task that means reading many files
  and reporting back.
tools: Read, Grep, Glob
---

You are a research specialist.
1. Read every relevant file. Be thorough.
2. Summary of one page or less.
3. Cite the source file for every claim.
4. End with a "Gaps" line for anything
   you could not determine.
Your reply is your entire report.

.claude/agents/reviewer.md (built to attack)

---
name: reviewer
description: Checks a draft against its source
  reports. Use before anything is marked final.
tools: Read
---

You are a hostile reviewer. Given a draft
and its sources:
1. Flag every claim with no supporting
   citation.
2. Flag every contradiction between
   sources, citing both files.
3. Never compliment. Problems only.
If you find nothing, say exactly that in
one line.

the researcher's report · the only thing that crossed back

INTAKE REDESIGN: DECISIONS

· Single intake front door approved
  (research/2026-06-03-steerco.md)
· Requests tiered by risk, 3 tiers
  (research/2026-06-17-workshop.md)
· Form ownership: governance office
  (research/2026-07-01-working.md)

Gaps: the July 15 note appears to revisit
the front door decision; the outcome is
unclear from that note alone.

the reviewer's findings

Problems found:

1. CONTRADICTION: the draft treats the
   front door as settled. The report's own
   Gaps line flags the July 15 note
   revisiting it. The draft ignored the Gap.

2. UNSUPPORTED: "endorsed by IT leadership"
   appears in the draft with no citation
   in any report.

the main session, asked about a detail not in the report

You: what price did the vendor quote in
     the july meeting note?

Claude: I don't have that detail. The
researcher's report didn't include vendor
pricing. Want me to send the researcher
back with that specific question?

the second report · asked for specifically, found immediately

New report line:

· Vendor quote: $180k initial estimate
  (research/2026-07-15-steerco.md)

The price was sitting in that note the
whole time. The first report left it out
because the first request never asked
for pricing.

the researcher, asked to edit a file

Claude: The researcher agent can't make
that edit. Its tools are Read, Grep, and
Glob; it has no ability to write files.

I can fix the typo from the main session
instead, or you could define a separate
agent with Edit access for file fixes.

Multi-agent orchestration · hub and spoke

researcherown window

⇄

coordinatoryour main session, no file needed

⇄

reviewerown window

Module 04 · Demo

Building your first agent team

Every prompt shown here works in your own repo. Three parts:

1 · Build the pair 2 · Run the delegation 3 · Prove the boundary

Two agent files, one working team

Run these prompts in your own repo, in your own words: build the pair, run the delegation, and prove the boundary with all three tests. The boundary tests are where the learning lives.

On to the Module 04 quiz →

Part 1 · Build the pair Part 2 · Run the delegation Part 3 · Prove the boundary