Agentic Coding
Agentic coding is software development in which an AI agent reads a codebase, edits files, runs tests, fixes failures, and returns a change for review.
Agentic coding is software development in which an AI coding agent carries out a programming task from start to finish: it reads the relevant code, plans a change, edits files, runs commands and tests, fixes whatever fails, and hands back a commit or pull request for review. The developer describes the task and judges the result instead of typing each line. Widely used coding agents include Anthropic's Claude Code, OpenAI's Codex, GitHub Copilot's coding agent, Cursor's agent mode, and open-source agents such as OpenHands, most of them driven by reasoning models trained to work through long chains of tool calls.
How Coding Agents Work#
A coding agent is an AI agent whose tools are a developer's tools: file search and reading, file editing, a shell for builds and tests, version control, and sometimes a browser. Its agent harness runs a loop in which a large language model (LLM) reads the task, inspects files, proposes an edit or a command, reads the output, and picks the next step. The design of those tools matters on its own: Princeton's SWE-agent resolved 12.5% of SWE-bench issues in 2024, far above earlier non-interactive methods, mainly by giving the model compact commands built for viewing, searching, and editing code, an approach its authors called the agent-computer interface.
Repositories steer agents with instruction files such as AGENTS.md, an open format for build commands, test commands, and coding conventions that the Agentic AI Foundation now stewards, along with agent skills that load task-specific procedures on demand and Model Context Protocol (MCP) servers that connect issue trackers, databases, and documentation. Agents run interactively in a terminal or IDE, or in cloud sandboxes where several tasks proceed in parallel and each one ends in a pull request. A lead agent can also hand searches, test runs, or reviews to subagents with their own context windows, a form of AI agent orchestration. Because a repository, a dependency, or a web page can carry a prompt injection, coding agents execute commands inside an agent sandbox and ask before pushing code, deleting files, or contacting new hosts.
Agentic Coding vs. Related Practices#
- Code completion: Suggests the next lines as a developer types, as GitHub Copilot did at its 2021 launch. The developer still writes and runs the program, accepting or rejecting each suggestion.
- Chat assistants: A chatbot answers programming questions with snippets that the developer copies, runs, and debugs by hand.
- Vibe coding: Andrej Karpathy's February 2025 name for building software by prompting and accepting AI-written code without reading it. It describes how a person treats the output, and agentic coding in professional teams usually keeps code review, tests, and continuous integration in the loop.
Agentic Coding in Computer Vision#
Machine learning operations (MLOps) work suits coding agents because each step ends in a checkable result: a dataset converts and validates, a training run starts and logs metrics, an exported model loads and returns the same predictions. An agent asked to fine-tune a detector can write the dataset YAML, launch training, read the validation metrics, and propose the next experiment. Ultralytics Agent Skills give Claude Code, Codex, Cursor, Gemini CLI, and other compatible agents this workflow as a yolo router skill plus skills for model selection, datasets, training, tuning, inference, and export, along with a platform-cli skill that drives Ultralytics Platform cloud training and deployments through the ul cloud command line. A request such as "export my trained YOLO model to TensorRT with INT8 quantization and benchmark it" then follows the export and benchmark procedures the skills document, and the skills tell the agent to trust the installed package's yolo checks and yolo cfg output wherever the two disagree.
Measuring Coding Agents#
SWE-bench, introduced in October 2023, is the standard benchmark for agentic coding: 2,294 real GitHub issues from 12 popular Python repositories, each scored by whether the agent's patch makes the project's own tests pass. The best model at release, Claude 2, resolved 1.96% of them. OpenAI's SWE-bench Verified, a 500-task subset that human engineers confirmed to be solvable, became the most reported version, and by late 2025 the best agents resolved about 80% of it. Harder suites such as Terminal-Bench test command-line work, and teams add their own agent evaluation on internal repositories. Productivity in practice is measured separately: in METR's randomized study of experienced open-source developers, 16 developers working on 246 tasks in mature projects they knew well in early 2025 took 19% longer when AI tools were allowed, mainly Cursor with Claude 3.5 and 3.7 Sonnet, while estimating afterward that the tools had made them 20% faster.










