Cursor: coding agents tutorial (2026)

Cursor’s agent harness combines instructions, tools, and a model. More specific prompts that refer to codebase patterns or include logs can improve results. Start a fresh conversation for a new feature or when the agent goes off track, and include known relevant files when useful.

Cursor agents use exact search, Instant Grep, and semantic search to find relevant code. The Explore subagent can search in its own context window and return findings without filling the main conversation with search details. For unfamiliar projects, ask for an architecture diagram, and understand existing patterns before asking the agent to change code.

Plan Mode helps research the codebase, clarify requirements, and create an editable implementation plan. In the example, the agent built model pages for a documentation site; the developer tested them locally, supplied a compile error, and used a screenshot to refine the design. For harder bugs, Debug Mode gathers runtime evidence with targeted logging, while basic debugging still depends on reproduction, isolation, hypotheses, and regression tests.

Treat agent-written code to the same review standards as other code. The examples include self-review with Find Issues, BugBot catching a clipped icon and duplicated function, and asking the agent to add unit tests. Agents can also test many user flows; one example ran for about 45 minutes and returned videos and screenshots.

Use short, shared rules for guidance needed in every conversation, and skills for specialized workflows loaded when relevant. Plugins can bundle rules, skills, MCP servers, and subagents, while agents can also use installed CLI tools such as GitHub CLI, AWS, Kubernetes, and Docker. The closing example applies the approach to checkout discount codes: first understand the system, then plan, build, debug, review, and record useful lessons in shared rules.