Agentic Coding

What Is Agentic Coding?

Agentic coding is the practice of producing software by directing an AI coding agent that reads the codebase, edits files, runs commands, and iterates on the results, instead of typing the code yourself. The developer supplies intent, constraints, and review; the agent supplies the keystrokes. The practice already operates at real scale: GitHub's Octoverse 2025 report counted over 1 million pull requests generated by its Copilot coding agent between May and September 2025, concentrated in established repositories rather than experimental projects [1].

Key Takeaways

• The developer's input shifts from code to instructions: task descriptions, acceptance criteria, and project conventions the agent must follow.
• The feedback loop is what makes it work. Agents that can run tests, builds, and linters correct themselves; agents that only generate text just guess with confidence.
• Repository context files such as AGENTS.md or CLAUDE.md are load-bearing. They tell every session how the project builds, tests, and expects code to look.
• Agentic coding is a practice, not a tool category. The same tool supports careful engineering or careless vibe coding depending on whether output gets reviewed.

How It Works

A session starts with a task given in natural language, ideally with acceptance criteria attached. The agent explores the repository, searching for relevant files and reading enough code to build a working picture of the change. Then it edits. After each meaningful edit it runs whatever verification the project offers: the type checker, the test suite, a build, a linter. Failures go straight back into its context, and it fixes them and reruns until the checks pass or it gets stuck and asks for help. Usage data reflects how much of the work the agent carries: in the Anthropic Economic Index analysis of 500,000 coding interactions, 79% of Claude Code conversations were classified as automation where the AI performs the task directly, versus 49% on the chat-based Claude.ai [2].

The quality of the outcome tracks the quality of three inputs. First, the task definition: vague requests produce plausible code that solves the wrong problem, which is why mature teams drift toward spec driven development. Second, project context: conventions written down in a context file are followed; conventions that live in someone's head are violated. Third, verification: a project with fast, trustworthy tests lets the agent close its own loop, while a project with no tests forces the human to be the only safety net.

Day to day, agentic coding changes the shape of a developer's attention. Less time in the editor, more time writing precise task descriptions and reading diffs. Review becomes the primary technical activity, since the diff is where wrong-but-plausible output gets caught before it ships.

Example

A developer needs pagination added to an admin endpoint. They write: "Add cursor-based pagination to GET /admin/orders, default page size 50, max 200, cursor must be opaque, update the OpenAPI spec, and extend the existing endpoint tests." The agent locates the handler, notices the codebase already has a cursor helper used by two other endpoints, reuses it, updates the spec, and adds four tests. The suite fails once because the agent's first cursor encoding broke on empty result sets; it reads the failure, fixes the edge case, and reruns clean. The developer reviews a focused diff and merges in under half an hour.

What People Get Wrong

The persistent misconception is that agentic coding means less engineering skill is required. What actually happens is that the skill relocates. Deciding what to build, decomposing it into agent-sized tasks, and critically reading generated diffs all demand the judgment of someone who could have written the code by hand. Teams that hand agents to developers who cannot evaluate the output do not get leverage; they get unreviewed risk merged faster, and comprehension debt that surfaces at the worst possible time. Perceived speed is also an unreliable guide: METR's 2025 randomized controlled trial with 16 experienced open-source developers found that using AI coding tools made them take 19% longer to complete tasks, even though they expected a 24% speedup [3].

FAQ

Is agentic coding the same as using Copilot-style autocomplete? No. Autocomplete suggests the next lines while you type and you remain the author of the file. In agentic coding the agent is the author: it navigates the repo, makes multi-file edits, and runs commands on its own, while you direct and review.

Does agentic coding work on large legacy codebases? Yes, often better than expected, because agents are good at reading unfamiliar code fast. The limiting factors are context: a codebase with no tests, no docs, and inconsistent conventions gives the agent nothing to verify against, so the human review burden goes up accordingly.

How is agentic coding different from agentic engineering? Agentic coding is the hands-on practice of one person driving an agent through tasks. Agentic engineering is the surrounding discipline: the specs, guardrails, and review processes a team builds so that practice produces reliable software at scale.

Sources

1. GitHub Octoverse 2025. "Copilot coding agent generated over 1 million pull requests between May and September 2025." https://github.blog/news-insights/octoverse/octoverse-a-new-developer-joins-github-every-second-as-ai-leads-typescript-to-1/. Accessed August 2026.
2. Anthropic Economic Index. "79% of Claude Code conversations classified as automation, versus 49% on Claude.ai, across 500,000 coding interactions." https://www.anthropic.com/research/impact-software-development. Accessed August 2026.
3. METR. "Randomized trial: experienced developers took 19% longer with AI coding tools despite expecting a 24% speedup." https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/. Accessed August 2026.