What is agentic development?

How agentic development works

Google Cloud defines agentic coding as "a software development approach where autonomous AI agents plan, write, test, and modify code with minimal human intervention".¹ What separates an agent from a chat assistant is the loop. Anthropic describes agents as "systems where LLMs dynamically direct their own processes and tool usage", which "are typically just LLMs using tools based on environmental feedback in a loop".² For a coding agent, the tools are the files, the terminal and the tests, and the feedback is what they report.

1. Takes a goal. A person describes the outcome they want: a bug fixed, a feature added, a dependency upgraded.
2. Plans. The agent reads the code it needs and breaks the task into steps.
3. Acts. It edits files and runs commands: installing, building, starting the app.
4. Checks. It runs the tests, reads what fails and fixes it, and goes round again until the task is done.
5. Hands back. The result comes back as a change to review: a branch, a pull request or a diff.

The person's work moves from typing the code to setting the goal and judging the result. That makes the review, the tests and the decision to merge matter more, not less.

It is still new to most teams. In Stack Overflow's 2026 developer survey, 83% of respondents used at least one AI tool at work, while agents and agent workflows reached 26%.³

How it differs from AI-assisted coding and vibe coding

Three ways of building with AI are often given the same name. They differ in who does the work and who checks it.

Andrej Karpathy named vibe coding in February 2025, for coding where you "fully give in to the vibes, embrace exponentials, and forget that the code even exists".⁴ Simon Willison drew the line a few weeks later: "When I talk about vibe coding I mean building software with an LLM without reviewing the code it writes." And: "If an LLM wrote the code for you, and you then reviewed it, tested it thoroughly and made sure you could explain how it works to someone else that’s not vibe coding, it’s software development."⁵ In Stack Overflow's 2025 survey, 72% of respondents said they were not vibe coding.⁶ Agentic development can be done either way. What makes it a way to build software a team will maintain is that somebody reviews and tests what the agent produced.

What teams use it for

Forrester's 2026 landscape of agentic development platforms names the work these tools take on: code autocompletion and generation, unit test generation and code explanation, and increasingly understanding large codebases, modernization, pull-request reviews and spec-driven development.⁷ In day-to-day terms:

• Fixing a bug from an issue report, with a test that shows it is fixed.
• Building a feature that touches several files.
• Writing the tests a codebase is missing.
• Upgrading dependencies and migrating old code.
• Explaining an unfamiliar codebase before changing it.
• Reviewing a pull request before a person does.

GitHub's agent shows the whole loop in one place: you assign it an issue, and Copilot "will start working on the task, raise a pull request, then request a review from you" when it is done.⁸

Tools for agentic development

Each of these is a coding agent: given a task, it does the work itself. They differ in where they run and how the work comes back.

Around the agents sit two more kinds of tools. An agentic development environment runs several agents at once and shows what each returns; an agentic development platform carries each change on through checks, merge and deployment. How they differ is on the platform explainer.

Risks, and what it needs around it

The evidence says reviewing what an agent wrote still matters. In Sonar's survey of over 1,100 developers, 96% do not fully trust AI-generated code, and only 48% always verify it before committing.¹⁵ New Relic found that 82% of technology leaders had at least one major production failure caused by AI code in the previous six months.¹⁶ Speed is not guaranteed either: in METR's study of 16 experienced open-source developers in early 2025, "when developers use AI tools, they take 19% longer than without", though METR now says those results no longer reflect today's tools.¹⁷

Agents also act on what they read. GitHub warns that people "can include hidden messages in issues assigned to Copilot cloud agent or comments left for Copilot cloud agent as a form of prompt injection", and requires that its agent's draft pull requests "must be reviewed and merged by a human".¹⁸ OWASP lists prompt injection first among the risks of applications built on language models.¹⁹

Google Cloud's advice is plain: "Before any code an AI agent makes goes into the main project, someone on the team should review it using the standard pull request process".¹ What teams put around the agents follows from that:

• An isolated workspace and branch for every task, so an agent touches only what it was given and several can work at once.
• The project's own tests and checks on every change, run automatically.
• A person who reviews the change and decides what merges.
• A way back: a deployment that can be rolled back to the last version that worked.