Agent plan mode

Agent plan mode is an interaction pattern, not a specific product feature. The idea: instead of letting an AI coding agent immediately edit files when you describe a task, you first get a structured plan — a list of steps, files to be changed, and the reasoning behind each decision. You can review that plan, push back on individual steps, or rewrite it entirely before the agent is allowed to touch anything. Only after you approve does execution begin.

This pattern emerged as a practical answer to one of the most common complaints about agentic coding: agents making changes in unexpected places, especially in larger codebases where a seemingly small instruction can cascade across dozens of files. Gemini CLI shipped an explicit plan mode in March 2026, restricting the agent to read-only tool calls during planning. Grok Build launched with plan mode as a headline feature in May 2026. Claude Code's subagents and hooks system gives teams similar control, even if the 'mode' framing is less explicit.

From a workflow perspective, plan mode pushes the human review surface up the stack: instead of reviewing individual diffs (line-by-line code changes) after the fact, you are reviewing intent before any code runs. Teams adopting spec-driven development (writing detailed feature specs before coding) find this pattern a natural fit, since the plan output functions like a machine-generated implementation spec you can sanity-check against the original intent. The pattern is still maturing, and different tools implement it differently, so it is worth checking how each tool handles plan approval, what happens when a step fails mid-execution, and whether the agent can revise its plan autonomously.