Skip to content

When work splits into parts that can go on at the same time, Mob proposes them as one plan: up to six coding tasks in one project, each for its own agent in its own worktree and branch. You approve the plan once, and every task starts.

Each row is one task: its title, its agent, what its blueprint runs after the agent’s work, and those steps as small pills (Build › Checks › Review › Ready). Approve it from the card, or open the flow first.

Open as flow, or click a row, draws the plan as a flowchart. The plan fans out into one lane per task: the agent builds, the project’s checks run, another agent reviews, and the task is ready. Dashed loops show what goes back to the agent, and how often: a failing check up to its retries, a review’s findings up to its rounds. Every lane ends with you, and shipping is always yours: Mob does not push or merge.

Click a step to see what it does. Details stay a click further: a task’s instructions sit under Instructions, and a started task’s last checks and the review’s conclusion under their own headings.

Studio, next to Flow at the top, draws the same plan as a studio of calves: one row per task, from the agent’s desk through the checks booth and the reviewer to your tray. Each calf does what its task is doing, and walks to the next station as the task moves on; a failed check sends it back to its desk with an “!”, and finished work lands in your tray. The number on a desk says where its task is, and opens that task’s steps in the flow.

While the plan waits for you, the step inspector edits the task:

Step What you can change
Agent the task’s instructions
Checks how many times a failing check goes back to the agent (0–3)
Review what it looks at (correctness, tests, hygiene, docs, security), how many rounds (1–3), or remove it
Ready Add a review before this, for a task without one

A changed step gets a dot, and the decision reads Approve with N changes; Undo changes puts the plan back as Mob asked. Mob checks your changes as it checks a blueprint before it starts anything.

A task of the manual blueprint leaves checking to Mob’s model, so it has no stages to change. Which agent works on a task cannot be changed here yet.

Once approved, each row shows where its task is, and its status opens the task’s conversation. The pills light up as the task moves: done in green, working in blue, waiting for you in amber.

In the flow, the step a task is on says what is happening, such as “Round 1” in review or “Fixing what the checks found”, and the inspector opens its conversation.

You hear once per task, when it is Ready to review or needs you (see Blueprints).

How it works

The tool. task_plan (owner only) takes a project, an optional summary, and one to six tasks, each with a title, a prompt, and optionally an agent and a blueprint (the project’s by default). It asks for one approval, then starts the tasks in order from the conversation that asked, as task_start would. A task that fails to start is reported and the others go on. The result names each started task’s place in the plan, so the card shows each start against its own task. Always allow covers plans in the same project with the same agents and blueprints. The manager prompt has the model propose work that splits into parts that can go on at once as one plan.

What the approval holds. The plan, with each blueprint as a line for people (does) and its stages for the card. Messaging channels show the tasks one numbered item at a time and one line per blueprint.

Approving with changes. An approval may carry a changed input (Answer.input). Only a tool that defines revise accepts one: it checks the change and returns the input to carry out, and the approval and the call keep what was approved. Other tools refuse a change and the approval keeps waiting. task_plan’s revise keeps the project, the tasks and their agents; it takes each task’s prompt and stages (null goes back to the blueprint’s), checks the stages with the blueprint rules (at most 3 retries and 3 rounds, each stage once, checks before review, known aspects), and drops stages that are the same as the blueprint’s.

A task’s own stages. A changed task starts with tasks.flow.stages, and Mob runs those in place of its blueprint’s (stagesOf in apps/mob/src/agents/blueprints.ts).

Where it lives. The tool and revisePlan are in apps/mob/src/agents/tools.ts; the card, the flow and the inspector in packages/app/src/PlanCard.tsx; the studio in PlanStudio.tsx and studioScene.ts, drawn with the landing page’s voxel calves (packages/ui/src/voxel/); approving with changes in packages/core/src/chat/conversation.ts (revise) and store.ts (decide, pendingApproval).