
Agentic coding: AI agents doing the coding, you owning it
In short: Agentic coding means AI agents plan, edit, run and test code across files while an engineer sets the goal, reviews the result and owns what ships.
Agentic coding is building software by having AI agents do the coding as multi-step tasks: reading the codebase, editing several files, running commands and tests, and iterating on failures. You set the goal, review the result and stay accountable for what ships. The SWE-bench paper framed the task in October 2023 as resolving real GitHub issues.
Agentic coding in plain words
You go from typing every line to being the tech lead of a fast junior who never gets tired, never gets bored and occasionally gets very confidently wrong.
How it works
- You describe a task with acceptance criteria.
- The agent explores the code, proposes a plan and makes changes.
- It runs tests, linters or the app, and fixes what fails.
- You review the diff, ask for changes, and merge.
Anthropic's Claude Code workflow guide follows the same shape, with recipes for exploring a codebase, fixing bugs, refactoring, running tests and opening pull requests, plus a plan mode that proposes changes before any edit touches disk.
My view: project instruction files, tests and fast feedback loops decide how well this works more than the model does.
Example
As an illustration: "Migrate the date utilities from moment to date-fns, keep behaviour identical, all tests green." The agent finds usages, replaces them file by file, runs the suite after each batch and reports two edge cases where behaviour differed for your decision.
Where you'll meet it
- Claude Code, OpenCode, Codex, Cursor's agent mode, Google Antigravity
- Background agents that open pull requests from issues
- Large mechanical changes: migrations, upgrades, test backfills
- Benchmarks such as SWE-bench, built from 2,294 real issues across 12 Python repositories
What it changes for you
For developers, value moves to problem framing, review and system knowledge, and writing tests first becomes a productivity tool. For tech leads, my view is that code review capacity becomes the bottleneck. I also expect junior roles to shift from writing boilerplate towards verifying, integrating and owning outcomes.
Agentic coding vs AI autocomplete
Autocomplete predicts the next few lines as you type. Agentic coding completes whole tasks across files and tools, then hands you the result. It also differs from vibe coding, which skips the review.
Common mistake
My view: measuring agentic coding by lines generated or tasks started. Measure accepted changes that stayed in production without rework. Everything else counts motion.
Related terms
Claude Code · Vibe coding · Spec-driven development · SWE-bench · Git worktrees for agents
Measure agentic coding by accepted changes that stay in production without rework.


