Agentic coding: when AI stops suggesting and starts doing
Agentic coding is a way of programming in which an AI agent takes on a whole task: it explores the repository, edits several files, runs the tests and fixes its own mistakes until it gets a result. The developer sets the goal, supervises and signs off.
Data as of October 22, 2026
Key figures
| Level | Advanced |
|---|
Definition
Put simply: instead of asking “write this function for me,” you say “fix this bug” and the AI handles it step by step. Agentic coding relies on an AI agent: a language model given tools (read a file, write one, run a command) through tool calling, looping between action and observation until the goal is met.
Anthropic, for example, describes Claude Code as an agentic coding tool that reads your codebase, edits files, runs commands and integrates with development tools, in the terminal, the IDE, a desktop app or the browser. Other tools follow the same logic, including GitHub Copilot, whose coding agent researches a repository, plans changes and implements them in the background.
A concrete example
A test starts failing after a dependency upgrade. The developer writes: “The billing test broke after the upgrade, find the cause and fix it.” The agent:
- searches for the relevant files and reads the error message;
- forms a hypothesis and edits two files;
- reruns the tests, hits a new failure, adjusts;
- presents the final diff and a commit message.
The developer reviews, asks for one tweak, then approves. Human effort shifts: less typing, more scoping and review.
Why it matters
Quality depends on the environment as much as on the model. The SWE-agent paper, posted on arXiv in May 2024, showed that well-designed “agent-computer interfaces” (commands tailored for navigating and editing code) significantly improve results. Hence the importance of context engineering: project instruction files, conventions, the tools exposed to the agent. Progress is tracked on tests such as SWE-bench.
Misconceptions and precautions
- “The agent replaces the developer.” It executes; responsibility for design, security and validation stays with people.
- “More autonomy is always better.” A vague task produces sprawling changes that are hard to review. Short, verifiable requests work better.
- “It compiles, so it’s fine.” An agent can work around a test instead of fixing the root cause. Review remains essential.
- Isolate execution: a sandbox and minimal permissions limit the damage from a wrong command or a malicious instruction hidden in a file.
Related terms: AI agent, tool calling, Terminal-Bench, context engineering.
Related terms
- AI agent · An AI that does more than answer: it plans and carries out a series of actions with tools to reach a goal you set.
- Tool calling (function calling) · The ability of an AI model to request that an external function be run with specific arguments, then use the result in its answer.
- SWE-bench · A benchmark that asks an AI to resolve real GitHub issues from Python projects, with each fix checked against the project's own tests.
- Terminal-Bench · A benchmark that measures whether an AI agent can complete hard, realistic tasks on its own inside a command-line terminal.
- Sandboxing · Running an AI agent inside an isolated environment with restricted files, network and commands, so any damage stays contained.
- Context engineering · The practice of deciding what goes into a model's context window at each step: instructions, tools, documents, memory and history.
Try it
Full glossary
Frequently asked questions
How is agentic coding different from code autocomplete?
Autocomplete suggests the rest of a line or function, and you decide everything. In agentic coding, the agent chains several actions on its own (searching, editing, running commands, rerunning tests) before handing control back to you.
Do I need to know how to code?
Not strictly to get a result, but it is strongly recommended to check that result. Without review, security or logic errors can slip through unnoticed.
Is it risky to let an agent run commands?
Yes, if it is not constrained. Good practice is to limit its permissions, run it in an isolated environment, require approval for destructive actions and review every change before merging.
How is a coding agent's quality measured?
With benchmarks that give the agent real issues to solve in real repositories, such as SWE-bench, or command-line tasks, such as Terminal-Bench. Scores change quickly from one model version to the next.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.