Start with the least powerful mode
A coding agent does not need write access to explain an error, map a module, summarize a diff, or propose a plan. Read-only work is the right default for unfamiliar repositories and tasks where the main risk is a wrong assumption.
Move into a mutating mode only when the task is clear, the relevant files are known, and you are ready to review the result. That keeps exploration useful without making every question an automation event.
Scope capability to the task
Permissions should distinguish reading, writing, terminal execution, Git actions, and network access. A broad 'allow everything' switch is convenient, but it removes the signal that helps a developer notice an unexpected request.
- Use Chat for non-mutating questions.
- Use Plan for read-only investigation and proposed steps.
- Use Agent when edits or commands are actually needed.
- Review requested actions when the policy asks.
Review is the final boundary
A permission prompt is not the last check. Inspect the changed files, read the Git diff, and run the validation that proves the task is complete. A clear stop button and visible activity make it possible to interrupt work before it expands.
Truss keeps these boundaries close to the agent workflow so a developer can grant the right capability for a real task without abandoning normal code review habits.
Next steps
Put it to work in your own workspace.
Pick the surface that fits your day, then connect a model you control.
FAQ
Questions developers ask before switching.
Are coding agents safe to use on production repositories?
They can be useful when work is scoped, permissions are explicit, sensitive actions are reviewed, and normal Git and test workflows remain in control.
What is the safest mode for a coding agent?
A read-only mode is safest for understanding a workspace. Grant writing or terminal capabilities only when a clear task requires them.