Key takeaways from Module 06: Agents and Role Specialization. Read this after the module; use it as your quick reference afterwards.
- Agent mode is the only mode with tool calls, multiple turns, and a working loop across files — and the most expensive mode to run incorrectly. It is correct only when all three conditions hold: multi-file/tool task, output not determinable in one turn, named deliverable with an exit condition.
- An agent persona applies Module 05's 4 design principles to a role, plus one critical addition: an exit condition. A persona without one runs until the context window fills and output degrades.
- Persona anatomy is exactly 4 fields: purpose, constraints, tool permissions, handoff criteria. A missing field makes the session unpredictable.
- The course ships a 10-role reference library in agents/. Study and adapt it — you author your own personas in the capstone, using templates/agent-definition-template.md.
- Roles that evaluate (Planner, Code Reviewer, Security Reviewer) never write to source files; implementation roles do. Evaluation produces text; implementation produces code.
- A session is a bounded execution loop with four phases — initialization → execution → completion → handoff — and context fills in one direction only.
- Higher context load means smaller scope: Security Reviewer, Solution Architect, and Performance Optimizer sessions must be scoped to one function or one small module.
- Open every session with the initialization prompt pattern: role, purpose, scope, constraints, tool permissions, exit condition — set at initialization, never mid-session.
- All 4 anatomy fields, always. Apply the completeness test: what does it produce, what may it not do, which tools can it use, how do I know it's done.
- Permission discipline: every ✅ must be load-bearing for the deliverable; evaluation roles get ❌ Edit — a reviewer that can edit is not a reviewer.
- One role per session — one scoped objective, ≤ 45 minutes of active turns. End the session the moment the deliverable exists.
- Every session names an exit condition and a completion signal; never continue past it just because the session still responds coherently.
- Hand off with the 3-part prompt — Summary, Objective, Carry-forward — and always reference
.github/copilot-instructions.mdin the carry-forward. - Pass committed file paths between sessions, not pasted output: the next role needs the task and its acceptance criterion, not the previous role's reasoning.
- Try auto model selection first; escalate to frontier only when multi-step causal reasoning demonstrably fails.
- Opening an agent session for a task a structured Ask prompt handles (5–15× the context for the same result).
- "Act as a senior developer" personas — no anatomy, no exit condition, no scope.
- Running Planner + Implementer + Reviewer in one session (context pollution).
- Granting Edit access to evaluation roles, or any tool "just in case."
- Duplicating instruction-file content in the init prompt instead of referencing it.
- Fighting session drift with more instructions instead of closing and re-initializing.
- Pinning a frontier model for every agent session.
| Role | Scope | Model tier |
|---|---|---|
| Planner / Analyst | Feature / epic | Auto |
| Solution Architect | Module / system | Frontier |
| Implementer / Developer | Function / module | Auto (frontier for complex algorithms) |
| Refactoring Specialist | File / module | Auto |
| Code Reviewer | PR / module | Auto |
| Security Reviewer | Function / file | Frontier |
| Test Engineer | Function / module | Auto |
| Documentation Writer | File / module | Auto |
| Performance Optimizer | Function / module | Frontier |
| DevOps / Release Assistant | Repo / pipeline | Auto |
Agent-mode test: multi-file or tools? · multi-turn? · named deliverable + exit condition? All three yes → Agent; otherwise Ask, Plan, or Inline.
Handoff prompt: Summary → Objective → Carry-forward. Reference structures for the common transitions: agents/handoff-prompts.md.
Before moving to Module 07, confirm:
- I can state the 3 conditions that make Agent mode the correct choice over Ask or Plan
- I apply the cost test — "could a structured Ask prompt produce the same result?" — before opening any agent session
- I can name all 4 anatomy fields and pass the completeness test on any persona I read
- I can name all 10 roles and state each one's primary output
- I know which roles have write access and why evaluation roles are always read-only
- I can fill in the permissions matrix for Planner and Code Reviewer without referencing the table, and justify every ❌
- I can state the 3 parts of a handoff prompt and write a Planner → Implementer handoff without a template
- I can name all 3 failure modes and identify which one is operating from a session description, quoting the evidence
- I know the session discipline rule — one role, one scoped objective, ≤ 45 minutes — and the 4 stop signals
- I know which roles justify a frontier model — Security Reviewer, and sometimes Solution Architect and Performance Optimizer — and why the rest run on auto
- I have read at least three definitions in agents/ and can adapt one to my own project without breaking the 4-field anatomy