Zed AI.

Agent works next to the code

Zed AI combines several ways to help in the editor: agent tasks, selection conversion, edit prediction and Git message preparation. For a big change, you can open an agent conversation, convey the context and observe the work with files. The resulting diff remains available to edit, accept or reject before the changes are fixed.

The editor has its own Zed Agent and connects external agents via ACP. This difference affects the settings, available models and powers. The external agent works through its process and configuration, so instructions to the built-in agent cannot be transferred to it without verification.

A good first task is a small mistake with clear expected behavior. The clearer the boundaries of change, the easier it is to evaluate the result and notice unnecessary changes. The common interface does not eliminate the need to read the code after the assistant.

Model selection and action management

For AI, you can use hosted Zed models, your own provider keys and local models. MCP tools extend the available context and actions. The documentation separately describes permissions, agent profiles and settings. A model with one set of features can work differently with tools than another, even with the same request.

Imagine a date-parsing function. Ask the agent to check an empty string, an invalid format and a month boundary, then propose a minimal fix. Request a specific input and expected result for each finding. This helps distinguish a real bug from a preference to rewrite the style.

After the changes, look at the diff and run appropriate checks on the project. If an agent offers a new dependency or changes the public interface, it is useful to understand separately the need for such a solution. This is an example of a problem statement, not a Zed personal test report or a promise to automatically fix all defects.

What is included in the free mode

Zed can be used for free, and AI connect with its keys; provider accounts remain separate. The placed models and their own prediction of edits have the terms of paid plans. The official reference describes the Pro trial period with a limited balance and a set of available features.

Local connection through Ollama allows you to use the model on your equipment. In this case, external agents and tools can access other services. For a closed project, check the entire configuration: where the code goes, what commands are available and which logs are saved. The speed of the interface and the quality of the generated change should be evaluated separately.

Suitable tasks

• Disassemble and fix a small part of the project
• Give the agent a limited task in the repository
• Convert selected code fragment
• Use local model in editor