Overview
Agentic coding workflows use LLM systems that autonomously plan tasks, write and edit code, run commands, and verify results. This goes beyond autocomplete into full task automation where the AI acts as an autonomous developer.
Plan-Execute-Verify Pattern
- •PLAN: LLM reads the codebase, understands the task, creates a step-by-step plan
- •EXECUTE: LLM implements the plan — edits files, creates new files, runs commands
- •VERIFY: LLM runs tests, checks types, reviews its own changes
- •ITERATE: If verification fails, LLM debugs and tries again
- •REVIEW: Human reviews only the final diff
Guardrails
File Restrictions
Limit which directories the agent can modify. Prevent changes to config files, CI pipelines, and production code.
Test Requirements
Require all tests to pass before the agent considers a task complete. This catches regressions early.
Iteration Limits
Set maximum retry attempts. If the agent cannot fix an issue in 3 tries, escalate to a human.
Tool Integration
Effective LLM agents need tools: file read/write, terminal commands, web search, and test runners. The richer the tool set, the more autonomous the agent can be. MCP servers, hooks, and custom commands extend agent capabilities.