AI Office

What it is and how to open it

AI Office opens as a full-screen window above the whole app. It is not one of the five pages of the main pager. It has no keyboard shortcut, no command-palette entry and no deep link: the four routes below are the only ones, and all four go through the same Pro check.

Ways to open it

Pro, checked once at the door

• AI Office is a Pro module as a whole. On Free, every route above shows the Pro paywall (the Office is listed under AI that works on its own, next to Cortex, Swarmcode and the AI Designer) and the Office does not open.
• Once you are inside, nothing else is gated by plan except the number of projects (see Free and Pro). AI providers are not part of the plan: you bring your own key.

How the window behaves

• Closing the window does not pause an office that is running. It keeps working, protected by a foreground notification, until it finishes or you pause it.
• Back (the system gesture): from Welcome it closes the Office; from My Offices, the Lobby and Activity it goes to Welcome (inside the Lobby it first goes back one page); from Analyze, Plan Review and AI Provider it returns to the Lobby; from Planning it cancels the planning and returns to the Lobby; from Board, Chat and Results it goes to Activity.

The team: eight roles

The Office has eight roles. They are not eight fixed agents: you choose the team, from one agent up to all eight, and there is one agent per role you select. The table says what each role does in the app.

Your team

• The Lobby lists the eight roles with a switch each. The Project Manager has a lock (Required) and can't be switched off, so a team has from 1 to 8 agents.
• The default team has five: Project Manager, Architect, Frontend Dev, Backend Dev and QA Engineer. Each template suggests its own team and always adds the Project Manager.
• Reviews exist only if the team has a QA Engineer: with none, tasks are accepted without a review.
• The Orchestrator you see in Chat and the project researcher behind Analyze are not roles: they are functions of the Office engine.

Which model each role uses

• You don't pick a model per role. When you press Start Office the Office chooses, for each role, the best model among the providers that have a key, and shows it on the role card of the team page, in the agent sheet of Activity and in Agent Performance in Results.
• It groups the roles in three levels: the strongest models for Project Manager, Architect and QA Engineer; fast and capable ones for Frontend Dev, Backend Dev and Data Engineer; the cheapest and fastest for DevOps and Tech Writer.
• If earlier sessions on the same linked project show another provider clearly doing better for a role (at least 3 samples and more than 15 points of success), it switches that role to that provider.

You can't, and why

• Choose the model, provider or reasoning level of a single role: the Office has no control for it, and the choice in AI Provider & Model doesn't set it either.Not yet
• Switch the Project Manager off: it is always in the team.By design
• Deploy from inside the Office: it is not connected to the Backend Manager, so there is no deploy step or button. Deploy from the Backend Manager.By design
• Use the Office with no AI provider connected: with no key, the first call fails and the Lobby shows “Planning failed — check the chat log for details”.Needs network

Welcome and My Offices

Welcome is the first screen of the Office. From here you open your offices or start a new session.

Welcome

Two actions and three information tiles under Features. In landscape the information tiles sit on the left and the actions on the right; in portrait everything is stacked in one column.

My Offices

The offices are the sessions you started: the current one and the past ones. Three counters on top: Total, Running (offices executing now) and Done (completed). A card shows the name, the state, the first lines of the brief, a bar with the tasks done, and chips with the number of agents, the number of sprints, the date, the estimated tokens (~Xk tok) and Project linked. The state has seven labels; an office waiting for a sprint review and a cancelled one are shown as Setup.

The Lobby: brief, team and Config

The Lobby is a two-page wizard. Page 1 is What should we build today?; page 2 is Your team, and the Config sheet opens from it. Back returns from page 2 to page 1.

Page 1: what to build

A big text box, three example chips, a carousel of templates and, if you already have projects, a way to build on one. “Describe it in your words. Your team handles the rest.”

Page 2: your team

Your team has the project name, one card per role and the Adjust details button.

The Config sheet

Adjust details opens a sheet titled Config. It exists only in the Lobby: once you press Start Office you can't change any of it. These are its controls, in order.

What the controls really do

• Quick Mode always skips the clarifying questions, asks the planner for 3 to 5 tasks of 2 to 4 files each and raises Parallel tasks to 2 if it was 1. It auto-approves the plan and skips the QA review only when the project is rated Trivial or Simple; from Moderate up you still review the plan and QA still reviews.
• Enable QA reviews: a review exists only if the team has a QA Engineer. If the review call itself fails, the task is accepted without review and the chat says so. On the last allowed retry, an output of more than 200 characters is approved automatically.
• Native tool-calling: on, the agents use the provider's native tool calls, with the same permission gates as AI Chat and the files jailed to the working folder. Off, they use the older text protocol, with a fixed list of tools per role and risk level and up to 8 tool rounds per task.
• Parallel tasks: the slider goes to 4 but the Office runs at most 3 tasks at a time, and only tasks the plan marks as parallel and that declare their files. The Feature Delivery playbook, SPARC and Auto (which defaults to Feature Delivery) raise the minimum to 2, which switches Best-of-N off.
• Max sprints: the planner is asked for at most the lower of Max sprints and the suggested sprints plus one. The suggestion is 1 for Trivial and Simple, 2 for Moderate, 3 for Complex and 5 for Massive.
• Best-of-N attempts: for each code task it makes N attempts one after another and keeps the one that passes the local check, so it costs about N times the tokens on those tasks. It only works with Parallel tasks at 1 and native tool-calling on, on the first attempt of each task.
• Adversarial review: a panel of N skeptical judges (1 to 5, with a quorum and up to 3 judges at once) reviews each task instead of the single QA Engineer. It needs QA reviews on and a QA Engineer in the team, and it multiplies the calls. Everything on this switch is shown in English.
• Design System, Playbook and Workflow playbook shape the prompts of the agents; they don't add tools or approval steps.

You can't, and why

• Choose SPARC: the Office has this playbook, but the selector shows eight cards and SPARC is not one of them. Auto turns it on only if your brief contains “sparc”, “specification pseudocode” or “pseudocode architecture”.Not yet
• Make Auto-approve plans do anything: nothing in the app reads it. To skip the plan review use Quick Mode, which works only for Trivial and Simple projects.Not yet
• Set a token limit, a time limit, a testing strategy or strict mode for tools: they exist in the Office's settings but no control shows them.Not yet
• Change the Config of an office after pressing Start Office: the sheet only exists in the Lobby.By design

Workflow playbook: how the work is planned

Each card tells the Project Manager how to plan and route the work: it shapes the planning prompt and which agent gets which task. It does not add tools or approval steps of its own.

Playbook: the quality bar

A playbook is a block of rules added to the prompt of every agent. Names and descriptions are fixed English text.

Planning and Plan Review

After Start Office the Project Manager plans. The Creating Your Plan screen shows three fixed lines (Assembling agents…, Analyzing requirements…, Creating task breakdown…) that are not a real progress bar, and a Cancel button; Back also cancels and returns to the Lobby.

What the planning does

1. 1It picks the workflow playbook (with Auto, from the words of your brief).
2. 2It asks the Project Manager's clarifying questions, unless Quick Mode is on.
3. 3It rates the complexity from 1 to 5: Trivial, Simple, Moderate, Complex or Massive. If the rating call fails it assumes Moderate.
4. 4It asks the model for the plan as sprints and tasks, each with a title, description, role, priority, dependencies, files, skills, acceptance criteria, risk, effort and whether it can run in parallel. The plan is written in your device's language.
5. 5It gives each task to the best-fitting agent of your team and, if the team has a QA Engineer, a QA Engineer other than the assignee as its reviewer.
6. 6It ends in Plan Review. With Quick Mode on and a project rated Trivial or Simple, it starts working straight away instead. If the plan can't be read, it creates a single catch-all task and says so in the chat.

Quick Questions

Unless Quick Mode is on, the Project Manager first asks 3 to 5 questions, each with 3 or 4 answer options, in your device's language. “Answer these to help the PM build a better plan.”

Review Plan

The screen shows the complexity (a chip in English: TRIVIAL, SIMPLE, MODERATE, COMPLEX or MASSIVE), the Strategy the Project Manager explains and the Sprints, each with its number of tasks.

Approvals and complexity

• Approving the plan is the Office's own confirmation. In projects rated Complex or Massive there is another one between sprints: the office stops in Sprint Review and the bottom bar shows Skip Sprint and Approve.
• Quick Mode skips the plan review only for projects rated Trivial or Simple. Auto-approve plans does nothing.
• From Moderate up, tasks estimated Large or Extra Large are split into smaller subtasks while the office works. Complex and Massive projects hold the code tasks back until the architecture tasks are done.

Activity: the office at work

Once the plan is approved the office works on four screens that you can swipe between or open with the four icons of the bottom bar: Activity, Board, Chat and Results. The header shows the project name, a dot that pulses while the office plans or works, and a line with the sprint, how many agents are working (N working) and the estimated tokens (~N tokens (est.)). The Save icon is on the right (see What it writes in your project). Activity is the home of a running office.

The action bar

Pause, resume and what they keep

• Pause stops at once: the call to the AI in progress is dropped and its partial result is lost, and a task in progress goes back to To Do and starts over when you resume. The files already written stay as they are.
• Resume relaunches the plan from where it was. Retry failed puts the failed tasks (with their retries reset) and the blocked ones back to To Do and resumes. Skip Sprint marks the sprint as skipped, blocks its unfinished tasks and moves on to the next one.
• When the work ends: if every task is done the office is Completed; if some are done and others failed or are blocked it stops Paused with a message such as “N/M tasks completed, N failed, N blocked”, so you can retry, re-plan or save what you have; if none succeeded it is Failed.
• A failed task is tried again up to Max retries per task times, waiting 3, 6, 12 and then 24 seconds before each new attempt. When two tasks in a row fail the office pauses by itself (“⚠ N consecutive failures detected. Pausing to reassess the plan…”). It also pauses when an agent asks you a question, and when its project folder is missing and your plan has no room for another project.

Tool approvals are decided in AI Chat

• The Office has no approval dialog of its own for tools. What the agents do with tools goes through the same gates as AI Chat: the tools you denied in AI Tool Permissions (Settings → AI), the check of how risky each terminal command is, the access level of the chat (No tools, Read only, Ask first, Auto-edit, Full access) and the approval of anything that leaves the device, such as a git push, a curl or an ssh.
• What needs your OK is parked in a queue and the agent keeps working: it is told that the action was not executed and was queued. The cards appear in AI Chat, and the Activity panel of the chat lists the AI Office under Needs you · N. Approve or deny there: the Office never asks inside its own screens.
• With the default level, Auto-edit, file edits run without asking, but running commands and anything that leaves the device wait for your OK. With Ask first, writing files asks too, also when the Office writes the code blocks of an answer at the end of a task: each file waits in the same queue and isn't written until you approve it. With Read only or No tools, the tools of the agents are refused and the Office does not write files by itself either.

What it writes in your project

This is the section that matters most: the Office writes into a real project of yours while it works.

Where it writes

• With a linked project, it works in that project's own folder and modifies its existing files; it is told to extend what exists and not to scaffold a new project.
• Without one, it creates a folder named like the project inside the app's projects/ folder, and that folder becomes one more project in your list. If a folder with that name already exists, the Office works in it. The name comes from the first three words of your brief unless you edit Project Name.
• At the start the chat says where it will write: “Working in the linked project: …” or “No project linked — the team will work in a new project: …”.
• The agents can't write outside that folder: paths with .. or absolute paths are dropped, and a symbolic link that points out of the folder is refused.

How the files are written

• While a task runs (native tool-calling, the default) the agents create and edit files and run commands with the tools of the app, jailed to the working folder.
• When a task finishes, the Office also writes the files found in the answer of the agent (code blocks with a path=, or the first code block into the target files of the task). With Ask first, each file goes through the same approval queue as the tools of the agents: it is parked in AI Chat and is not written until you approve it. With Auto-edit or more it is written without asking; with the access level Read only or No tools this step is skipped and the Office writes nothing by itself.
• A QA review reads text. On top of it the Office checks the work locally: the HTML and CSS of the Frontend Dev get a structure check, and a script the agent wrote (Python, JavaScript, Ruby, PHP and some more) is run on your device with the same runners as the Run button of the editor, for 30 seconds at most. A real failure sends the task back for another attempt (first attempt only, with QA reviews on).
• A separate honesty check compares what a report says it ran with the commands that were really executed. A task that boasts of a test run nobody launched is sent back or, with no retries left (or when the agent has no way to run commands), finishes marked Not verified by PocketCode.

What it asks you

• Your approval of the plan, and of each sprint in Complex and Massive projects. It is the only confirmation of the Office itself.
• The data notice of your AI provider, once per provider and account: the same notice and the same record as AI Chat. If you decline, nothing is sent.
• Tool approvals, which are decided in AI Chat (see Activity).

The Save icon

• Save, in the header from Plan Review on, writes a summary and the files of the task results. In a linked project the summary goes to .pocketcode/office/session-<id>.md, so that it doesn't overwrite your README.md; otherwise it is the README.md of the project. Afterwards the icon turns into a green check and no message is shown.
• You don't need it to keep the files: they are already in the project (under Ask first, apart from those still waiting for your approval).
• Save writes the code the agents left in the task results, never their report on top of a file: a block with a path= goes to that path, the first code block of a task goes into its target files, and a result with no code block is written to tasks/<title>.md, so the target files of the task are not touched.

Pause, close and recovery

• Every finished task is saved to the device, so an office survives a restart with its finished work.
• If the app closes while an office runs, open it from My Offices: it comes back paused, with the message “This office was interrupted when the app closed (X/Y tasks done). It's paused — tap Resume to continue.”, and you press Resume.
• After each sprint the Office also stores a checkpoint of the session, but nothing in the app restores from it.

You can't, and why

• See a diff, accept or reject a file, undo what the Office wrote or list the files it touched: the engine keeps a private record of which agent changed what, but no screen shows it.Not yet
• Restore an office from its checkpoints: they are saved after each sprint, but nothing in the app uses them. Recovery works through My Offices and Resume.Not yet
• Make an agent write outside the working folder: the paths are checked against the real location of the folder.By design

Privacy and cost

AI Office sends the content of your project to the AI provider you chose, straight from your device: PocketCode puts no AI service in between. Treat the brief and the files you let the agents read as things your provider will process.

What goes to the provider

• Your brief, your answers to the questions, the plan and the tasks, in the calls that need them.
• The context of each task: the earlier results (each cut to 1,500 characters), the files the agent reads with its tools and what its commands print and, with a linked project, its path, its stack and up to 60 file names.
• The instructions file of the project, if it has one (its first 4,000 characters).
• With Analyze: the manifests and docs of the project plus up to 17 source files, each cut to 2,000 characters.
• Lessons and memories: up to 10 lessons from the session, up to 8 relevant memories per task and up to 12 memories from the last 5 sessions of the same linked project.
• The first time you use each provider, the Office needs your acceptance of its data notice, the same one AI Chat shows. It is asked once per provider and account, and nothing is sent to a provider you haven't accepted.
• The flag icon on AI messages, artifacts and post-mortems opens Report response: you choose a reason (child sexual abuse or exploitation, hate or harassment, violence, sexual content, illegal or harmful instructions, or something else) and Send report sends it to PocketCode.

What stays on your device

• Each office is saved in the private storage of the app after every finished task: the brief, the plan and the tasks with their results, the last 200 chat messages, the memories, the artifacts, the Config and the path of the linked project. Deleting an office from My Offices removes it from the list and from the database.
• The sprint checkpoints are JSON files in the private storage of the app.

What it costs

AI Office spends the quota or the bill of your own provider. PocketCode sells no AI credits and shows no amount of money: the only numbers you see are token counts, always an estimate (~N tokens (est.)).

• You can't set a token or time limit. What bounds the spending: Pause and Cancel Session, Max retries per task, Max sprints, a smaller team, Quick Mode on simple projects and the automatic pause after two failed tasks in a row.
• Every call has a 4-minute timeout. After a failure it is tried once more and then with another connected provider, unless the problem is a missing data notice.
• The estimate shows in the header, in the agent sheet, in the team strip (~1.2k), in Agent Performance and on the card of each office in My Offices (~Xk tok).

The calls an office makes