agentic-engineering

Agentic Engineering

Design tasks so AI agents can execute them reliably. This is not about using Copilot features — it is about architecting work so agents succeed on the first attempt, fail loudly when they can't, and hand off cleanly to the next agent.

Why This is Copilot-Exclusive

The patterns here are specific to Copilot CLI's agent execution model: task() dispatch, read_agent / write_agent lifecycle, SQL session state, and background agents with mode: "background". They don't map directly to interactive coding sessions in other tools.

When to Use

• Decomposing a large task before dispatching it to an agent or fleet
• Designing a multi-agent workflow where context must transfer between agents
• Debugging why an agent produced incorrect or incomplete output
• Establishing quality standards for a new agentic workflow

When NOT to Use

Instead of agentic-engineering	Use
You already have tasks and just need to plan them	plan-mode-mastery
You need to assemble a specialist agent team	team-planner
You need autonomous execution guardrails	autopilot-patterns

Core Principles

1. The 15-Minute Task Unit

Rule: Each agent dispatch should complete in roughly 15 minutes of human-equivalent focused work. In practice: 1–3 files changed, 1 clear outcome, no more than one decision required.

Why: Agents fail when context exceeds what fits in a single focused pass. Long tasks require the agent to hold too much state, make too many decisions, and produce outputs that are hard to verify.

Signs a task is too large:

• Description contains "and" more than twice
• Requires reading more than 5 files to complete
• Has more than one possible success state
• Cannot be verified by a single test or check

Signs a task is too small:

• It is just a file read or a lookup
• A single edit call handles it entirely
• No judgment is required

Decomposition pattern:

Large task: "Implement user authentication with JWT and refresh tokens"
↓ decompose
T-01: Add User schema + bcrypt password field (DB layer only)
T-02: Implement POST /auth/login endpoint (validate + sign JWT)
T-03: Implement POST /auth/refresh endpoint (validate refresh token)
T-04: Add auth middleware (extract + verify JWT on protected routes)
T-05: Integration tests for T-02, T-03, T-04