Manus adds Plan Mode, drafting specs before agents execute
The agent startup's new Plan Mode turns rough prompts into reviewable specs before execution, bringing a pattern proven in coding agents to general tasks.
Manus introduced Plan Mode on July 22, a feature that has the general-purpose agent draft a structured plan before it starts executing. Instead of demanding a carefully engineered prompt up front, the system takes a rough idea, expands it into a clear specification, and lets the user review and adjust the direction before any work begins. The company framed the release around a simple claim: users should not need a perfect prompt to get a good result.
Plan first, build second
In practice, Plan Mode inserts a checkpoint between intent and autonomy. The agent proposes what it understands the task to be, how it intends to approach it, and what the deliverable will look like; the user confirms or redirects, then execution proceeds against the agreed spec. The pattern will be familiar to developers — coding agents such as Claude Code popularized an explicit planning mode in which the model must present its approach before touching files — but Manus is applying it to the much broader surface of research reports, websites, slides, and data tasks its agent handles.
The launch continues an aggressive July shipping cadence for the company, following Branch (parallel exploration from one context) on July 9, Auto-Publish on July 13, and native PowerPoint generation on July 14.
Why it matters: the biggest hidden cost of autonomous agents is a long run heading in the wrong direction — wasted minutes, wasted credits, and eroded trust. A mandatory plan-review step attacks that failure mode directly, and its spread from coding tools to consumer-facing agents suggests spec-first interaction is hardening into standard agent UX. As the underlying models converge, agent products are increasingly competing on exactly this layer: workflow control, checkpoints, and the ergonomics of steering, rather than raw capability.