Sub-Agents + Build Tools

Level 4 · Course 24

More agents do not automatically mean more progress. They can multiply output, but they can also multiply mistakes. A vague prompt sent to three helpers gives you three confident wrong answers. A build tool without a scoped handoff can change files you did not know were in play.

This course teaches your agent when to bring in help and how to keep that help contained. Sub-agents are for controlled evaluation, research, and clean-room checks. Build tools are for scoped implementation. Each helper gets a role, inputs, allowed files, forbidden actions, expected output, and stop condition.

You stay in command. Your main agent manages the handoff, reviews the result, and does not treat helper output as truth just because it came from another model. After this course, extra agents become a controlled panel or construction crew — not a swarm.

Example One: QA Before Launch

A marketing operator needs to QA a landing page before it goes live. The main agent drafted the copy. Instead of reviewing its own work — which compounds errors rather than catching them — it hands off to a sub-agent.

ROLE: Adversarial copy reviewer

INPUTS:
  - Landing page HTML (attached)
  - Target audience profile (attached)
  - Brand voice guide (attached)

ALLOWED FILES:
  - Read: landing-page.html
  - Read: brand-voice-guide.md

FORBIDDEN ACTIONS:
  - Cannot edit landing-page.html
  - Cannot access client data
  - Cannot send anything externally

EXPECTED OUTPUT:
  List of 3–5 specific issues, ranked by severity.
  Each issue: what it is, why it matters, suggested fix.

STOP CONDITION:
  Return after one pass. Do not iterate.

The sub-agent returns its list. The main agent passes it to the operator. The operator decides which fixes to make. The sub-agent does not touch the file. That separation is the point.

Example Two: Research Handoff

The same operator needs competitive intelligence before a pitch. Rather than running a long research thread through the main session, they hand off to a scoped research sub-agent.

ROLE: Market researcher

INPUTS:
  - Competitor URLs (listed inline)
  - Industry keywords (listed inline)

ALLOWED FILES:
  - Web access only

FORBIDDEN ACTIONS:
  - No file writes
  - No external communication beyond reading the listed URLs

EXPECTED OUTPUT:
  One-page brief. Top 3 findings with sources.
  No editorializing. Facts and citations only.

STOP CONDITION:
  5 minutes or 3 findings — whichever comes first.

The stop condition matters. Without it, a research agent runs indefinitely, pulling tangential sources and inflating the output. The constraint forces prioritization.

When Not to Use Sub-Agents

Simple one-step tasks. If the main agent can answer in one pass, a sub-agent adds overhead with no benefit.

Anything requiring operator judgment. Sub-agents return output, not decisions. If the question is "should we run this campaign," that answer belongs to the operator, not a helper.

Anything where the main agent already has the answer. Handing off a task you could complete in the current context wastes time and introduces a round-trip that can distort the original question.

You decide: which task do you hand off first? Pick something you currently review yourself that does not require your judgment — just your verification.