Agent lesson 02 · Delegate with contracts
Shared teaching scenario · fictional product data
Research three laptops under €900, verify current evidence, and write laptop-comparison.md. Do not purchase anything or contact a vendor.
01 · Smallest useful mechanism
Parallel workers are useful when work can be partitioned without hiding critical dependencies. Each delegation needs an explicit objective, context packet, allowed tools, output schema, budget, and completion condition. The parent compares and reconciles results rather than trusting an appealing summary.
A subagent should receive a bounded question and return an inspectable artifact; the parent remains responsible for integration and the final claim.
02 · Experiment
Deterministic browser simulation
The delegate with contracts experiment uses inspectable, repeatable teaching data. It does not claim to be a live agent trajectory.
03 · Production-minded practice
Use your own stack or a contained mock environment. Do not point failure drills at real accounts, people, or irreversible services.
Write contracts for a specification researcher, review researcher, and evidence auditor, including inputs, tools, budgets, output schemas, and stop conditions.
Return conflicting weight measurements from two workers and let a third worker time out with a partial artifact.
The parent preserves both claims, requests only the missing evidence, resolves or discloses the conflict, and remains the owner of the final recommendation.
Your learning artifact
Stored only in this browser. No account required; course reset does not delete it.
04 · Check your understanding
Next: Every worker and every turn needs a carefully constructed working set, so the next lesson opens the context harness.
Delegation design
Find three candidate laptops
9 min · 2,800 tokens
read-only investigation
Verify prices and specifications
11 min · 3,900 tokens
read-only investigation
Challenge missing or stale evidence
7 min · 2,400 tokens
read-only investigation
Elapsed time
11 min
Total child tokens
9,100
Parent context
1,250
No shared-write conflict in this task split. The parent can review outcomes before integrating them.
Research basis: Codex subagents, Claude Code subagents, and worktree isolation.