Sequential Thinking

Sequential Thinking gives an agent a dedicated tool for externalizing multi-step reasoning instead of trying to hold an entire plan in one pass. Each call records a discrete “thought,” and the tool supports revising an earlier thought, branching into an alternative line of reasoning, and expanding the total number of steps mid-task as the scope of a problem becomes clearer.

It’s most useful on problems where the right approach genuinely isn’t obvious at the start — debugging something with several plausible causes, designing an approach before writing code, or working through a task that may need course-correction partway through. Rather than committing to a single linear chain of thought, the agent can mark a step as a revision of an earlier one or fork into a new branch, which keeps the reasoning trace honest about where it changed direction.

There’s no external state to configure — the only setting is whether intermediate thoughts are logged, which you can disable in environments where you don’t want that trace persisted.