1. The agent receives a task (a feature description, a bug report, a command) and access to the repository. 2. It plans steps and searches the code to locate relevant files. 3. Through an agent-computer interface it takes actions: editing files, running commands, building, testing. 4. It observes the outcome (logs, errors, test results) and feeds it into the next reasoning iteration. 5. It repeats the plan โ act โ observe cycle until tests pass or the task is complete. 6. It presents the changes (e.g. a diff / pull request) for review. The foundation is an LLM with tool-use capability, often embedded in a harness that manages context, memory, and permissions.
Classic code assistants (autocomplete, editor suggestions) require the developer to drive the whole process: understand the repository, run tests, and integrate changes. A coding agent automates these multi-step, tedious engineering tasks โ from understanding an issue to fixing it, running tests, and preparing changes โ reducing repetitive manual work and shortening turnaround.
The set of tools exposed to the model (file editing, code search, running commands and tests), designed for reliable agent operation.
An iterative loop where the agent plans, takes an action, observes the result, and adjusts subsequent steps.
Infrastructure managing context, memory, permissions, and tool-call orchestration around the LLM.
The agent's multi-step operation produces long trajectories that exceed the LLM context window.
An agent with terminal and file access can execute harmful commands or be manipulated by repository/web content.
Large, autonomously produced changes can be hard to assess, complicating correctness verification.
Mass adoption of LLM-based code assistants as in-editor suggestions and autocomplete.
Established the standard for measuring models'/agents' ability to solve real software engineering tasks.
Agent-computer interfaces and full agent loops enabled autonomous task solving; commercial products appeared (e.g. Devin).
Coding agents became everyday tools, with SWE-bench Verified scores surpassing 70%.
