Exercise 06: Extend Pipeline

Objective

Add a new team-linter agent to the pipeline. This tests understanding of the agent extension mechanism.

Required Reading

• Practitioner README -- "Extending the setup > Add a new agent" section
• Customizing Agents | Cursor Learn -- How to create and register custom agents
• Custom Agents | Cursor Docs -- Agent .md frontmatter (name, model, description), AGENTS.md registry format, and subagent_type mapping

CursorClaude Code

Adding a new agent in Cursor uses AGENTS.md with subagent_type mapping.

In Claude Code, you define the agent's role, model, and responsibilities. The difference is the dispatch mechanism: direct model invocation. The agent definition (role, responsibilities, output schema) is the transferable design artifact.

Context

The pipeline works end-to-end (Exercises 01-05 complete). Now extend it with a team-specific agent.

Tasks

1. Copy the agent template:

   cp sandbox/.cursor/templates/agent.md sandbox/.cursor/agents/team-linter.md
   

2. Edit sandbox/.cursor/agents/team-linter.md frontmatter:

   ---
   name: team-linter
   model: gemini-3-flash
   description: Runs project linter and writes lint result; use when verifying code style before tests.
   ---
   

3. Fill in the body:

4. ROLE: Runs the project linter and writes a lint-result.json artifact
5. CORE RESPONSIBILITIES: Read plan/worker-result, run npm run lint, write .pipeline/<issue-id>/lint-result.json with verdict and output
6. NON-GOALS: Does not fix lint errors (that's the worker's job)

7. OUTPUT: lint-result.json with { "verdict": "PASS"|"FAIL", "output": "...", "errors": [] }

8. Add team-linter to sandbox/.cursor/AGENTS.md:

9. New row in the agents table
10. Pipeline order note: "2.5. team-linter -- After worker; writes lint-result.json. On FAIL -> planner re-dispatches worker."
11. Subagent types: add linter -> team-linter

Validation

python3 docs/practitioner/tutorials/verify.py --exercise 06

Checks: team-linter.md exists with valid frontmatter, AGENTS.md references team-linter and lint-result.json.

Reflection

• How does adding a new agent affect the pipeline flow?
• What would you need to change in the planner to dispatch to the team-linter?
• Could this agent be tiered (fast/standard/high)? When would that make sense?

Answer

team-linter.md frontmatter:

---
name: team-linter
model: gemini-3-flash
description: Runs project linter and writes lint result; use when verifying code style before tests.
---

Body should define: ROLE (runs linter, writes artifact), CORE RESPONSIBILITIES (read plan/worker-result, run npm run lint, write lint-result.json), NON-GOALS (does not fix lint errors), OUTPUT schema (verdict, output, errors).

AGENTS.md additions: new row in table, pipeline order "3.5. team-linter", subagent type linter -> team-linter.