What is the difference between an AI coding assistant and an AI coding agent?

An **AI coding assistant** primarily helps a developer decide or write: it answers questions, explains code, suggests completions, drafts functions, or proposes changes.\n\nAn **AI coding agent** can carry out a multi-step engineering task within a defined environment. It can inspect a repository, plan work, edit multiple files, run commands and tests, observe failures, revise its approach, and return a validated patch or pull request.\n\nThe distinction is not binary. Products exist on a spectrum of autonomy. The useful test is not what the product calls itself, but **what actions it can perform, what feedback it can observe, how long it can continue, and what authority it has**.

A conventional assistant operates mainly in a request-response loop. The developer remains responsible for selecting files, applying suggestions, running tests, interpreting failures, and deciding what to do next. The assistant may be highly capable, but it does not independently close the loop between suggestion, execution, observation, and correction.\n\nA coding agent closes more of that loop. A typical agent can:\n\n1. read repository files and project instructions;\n2. search for relevant symbols and dependencies;\n3. create a plan;\n4. modify one or more files;\n5. invoke a terminal, compiler, linter, test runner, or browser;\n6. inspect the result;\n7. revise the implementation after failures;\n8. preserve a trace of its actions;\n9. commit changes or propose a pull request for human review.\n\nOpenAI's description of Codex illustrates this difference: tasks run in isolated repository environments, Codex can read and edit files, run commands including tests and linters, iterate until tests pass, commit changes, and provide terminal evidence for review.\n\nAutonomy does not remove the need for governance. A coding agent that can run commands or access networks has a larger failure and security surface than an assistant that only generates text. The organisation must define:\n\n- readable and writable paths;\n- network destinations;\n- secret and credential access;\n- allowed commands and tools;\n- approval boundaries;\n- maximum cost and runtime;\n- test and review requirements;\n- whether it may commit, open a pull request, merge, deploy, or publish.\n\n“Agent” should therefore describe an execution capability, not a guarantee of competence. A product may be agentic yet unreliable, unsafe, or unable to complete the intended class of work. Evaluation should measure end-to-end task completion under realistic constraints.\n\nA useful vocabulary is:\n\n- **assistant** — advises or generates;\n- **agent** — observes, acts, and iterates toward a goal;\n- **workflow** — follows a mostly predefined sequence;\n- **autonomous agent** — selects more of its own intermediate steps;\n- **governed agent** — acts through enforceable permission, approval, and audit boundaries.

- [OpenAI — Introducing Codex](https://openai.com/index/introducing-codex/) — describes a software-engineering agent that reads and edits files, runs commands and tests, iterates, commits changes, and proposes pull requests. - [OpenAI — Running Codex safely](https://openai.com/index/running-codex-safely/) — establishes that coding agents autonomously review repositories, run commands, and interact with development tools, requiring execution boundaries and approvals. - [Anthropic — Claude Sonnet 5](https://www.anthropic.com/news/claude-sonnet-5) — describes an agentic model that plans, uses browsers and terminals, and operates autonomously. - [Model Context Protocol — Architecture overview](https://modelcontextprotocol.io/docs/learn/architecture) — defines standard mechanisms through which AI hosts can discover and invoke tools and retrieve resources.