What is agentic coding?

Agentic coding is a way of developing software in which an AI agent takes a task, then reads code, edits files, runs commands and checks the result in a loop, with you setting the goal and reviewing the outcome. Simon Willison defines the underlying idea as “agents run tools in a loop to achieve a goal”, and uses “agentic engineering” for “developing software with the assistance of coding agents”. It is the coding case of agentic AI, and a coding agent is one kind of AI agent.

How the loop works

Anthropic’s Claude Code docs say a task goes through three phases: “gather context, take action, and verify results”. The model decides each step from what the previous step returned, “chaining dozens of actions together”, and you can interrupt at any point (Claude Code docs).

Tools are what separate this from a chatbot. The same docs say that “without tools, Claude can only respond with text.” The built-in tools fall into five groups: file operations, search, execution (shell commands, servers, tests, git), web, and code intelligence. The software around the model that supplies the tools and manages its context is called the harness (Claude Code docs).

Other products describe the same shape. OpenAI says Codex CLI lets you “inspect code, make changes, run commands, and automate repeatable work” from the terminal. GitHub says its Copilot coding agent “can research a repository, plan changes, and implement them in the background”, then lets you review before a pull request is created (GitHub Docs).

Agentic coding compared with nearby terms

“Agentic coding” is used for the middle and right columns, and sources do not agree on one definition. IBM treats vibe coding as the informal end of a spectrum and agentic coding as adding “defined roles and constraints”. Fowler and Willison draw a harder line at whether anyone reads the code. For the fixed-path alternative, where your code picks the steps, see agentic workflows.

What you control

Most tools expose the same controls.

• Permissions. Claude Code has Manual, Accept edits, Plan and Auto modes; Plan “explores and proposes a plan without editing your source files” (Claude Code docs). Codex has a /permissions command to choose when it can edit files or run commands without asking (OpenAI).
• Undo. Claude Code snapshots a file before editing it, but “actions that affect remote systems (databases, APIs, deployments) can’t be checkpointed” (Claude Code docs).
• Persistent instructions. Each session starts with a fresh context window, so rules that must always apply go in a file such as CLAUDE.md rather than in the conversation (Claude Code docs).
• Isolation. MIT’s Missing Semester warns that “yolo mode”, where the agent runs tools without asking, is not secure to use directly and suggests a virtual machine or container (MIT Missing Semester). GitHub runs its agent in a GitHub Actions runner or cloud sandbox (GitHub Docs).

Verifying agent output

Agent output can look right and still be wrong. IBM warns that output can “look correct but contain subtle bugs or security vulnerabilities” and says to treat it as “a first draft that requires careful review”. GitHub says Copilot “can produce incorrect or suboptimal code, including code that contains security vulnerabilities” (GitHub Docs).

Anthropic and MIT both advise giving the agent something to check against:

• MIT: set things up “so that the model can run the failing check directly, which will let it iterate autonomously” (MIT Missing Semester).
• Anthropic lists “tests, typecheck, a Stop hook, CI” as ways to define when work is done (Claude Code best practices).
• Anthropic also suggests a reviewer in a fresh subagent that “never saw your conversation”, so it “isn’t swayed by the reasoning that produced the code” (Claude Code best practices).
• MIT adds a limit: “sometimes verifying code can be harder than writing the code yourself; for critical code, consider writing it by hand”, and warns of “debugging spirals” (MIT Missing Semester).

Benchmarks such as SWE-bench score how well agents resolve real repository issues. Agents that read untrusted text, such as issues or web pages, are also exposed to prompt injection, and guardrails limit what a hijacked agent can do.

Does it make developers faster?

The best controlled evidence is from METR, and it has moved. In a 2025 randomized trial, 16 experienced open-source developers worked on 246 real issues, each randomly assigned to allow or disallow AI. Developers took 19% longer with AI, although they expected a 24% speedup and afterward believed AI had sped them up by 20% (METR, 10 July 2025). METR now marks those results as out of date.

A larger follow-up covered 57 developers and over 800 tasks. It estimated a speedup of 18% for the original developers (confidence interval from a 38% speedup to a 9% slowdown) and 4% for new recruits (15% speedup to 9% slowdown). METR writes these as changes in task time, “-18%” and “-4%”, so a negative number means faster. METR says the estimate is likely a poor measure, because 30% to 50% of developers withheld tasks they did not want to do without AI. It expects developers are now more sped up than in early 2025, but calls its data “very weak evidence for the size” of that change (METR, 24 February 2026). Neither study gives a reliable number for current tools.

Seeing what a coding agent did

An agent session is a trace of prompts, tool calls and results, so reviewing it works like reviewing any agent run. HoneyHive’s Claude Code integration exports session traces with each prompt and response, built-in tool calls (Bash, Edit, Read, Grep) with arguments, results and duration, and reasoning blocks (HoneyHive docs). Setup is pip install honeyhive-daemon, then honeyhive-daemon init in your repo, an HH_API_KEY, and honeyhive-daemon run. It installs Claude Code hooks at the user level, so your code does not change. A separate Codex CLI integration sends Codex’s native OpenTelemetry traces.

FAQ

What is the difference between agentic coding and vibe coding?

In vibe coding you do not look at the code. Martin Fowler writes that "with vibe coding humans don't look at the code, indeed they forget that it even exists", while in agentic engineering they "are concerned with the code, often giving it detailed review". Simon Willison reserves vibe coding for "unreviewed, prototype-quality LLM-generated code".

How is agentic coding different from autocomplete?

Autocomplete suggests the next line from nearby code. IBM says earlier assistants "analyzed existing syntax and nearby code to suggest the next line of code". A coding agent also reads the repository, runs commands and uses the results to decide its next step (Claude Code docs).

Does agentic coding make developers faster?

The evidence is mixed and dated. METR's 2025 trial found experienced developers took 19% longer with AI tools, and METR now calls those results out of date. Its February 2026 follow-up says developers are probably more sped up now, but the data is "very weak evidence for the size" of that change (METR).

Is it safe to let a coding agent run commands?

Only with limits. MIT's Missing Semester says "yolo mode", where the agent runs tools without asking, is not secure to use directly and advises running the agent "in an isolated environment like a virtual machine or container" (MIT Missing Semester). See also prompt injection.

How do I see what a coding agent did?

Record each session with its prompts, tool calls and results. HoneyHive can export Claude Code session traces, and see agent observability for what a trace should hold.