Agent & tasks · 14 min read · Updated 2026-07-30
Ask, Run, Plan, or PMO: which mode should I use?
What each mode can do, where to switch, when to use it, and what happens if you choose the wrong one.
Give Agent the same request and it might explain the issue—or immediately edit your files. The difference is themode—the small label beneath the chat box.
Mode determines how far Agent may go in the next turn: research and answer only, draft a plan first, start working directly, or act as a project manager and coordinate a team of sub-Agents.
The right mode removes friction. The wrong one either leaves Agent unable to act or lets it change a lot before you have settled the direction.
Where to switch modes
- Open any Agent conversation
- Find therow of small buttons below the composer,The leftmost buttonis the mode selector
- It shows the current mode's icon and name (the default is Run), with a dropdown arrow on the right
- Click it to open a menu. The current mode has a checkmark, and hovering over an item shows a short description
- Click once to switch; there is no confirmation dialog—when the label changes, the mode has changed
You can also use the shortcut Cmd/Ctrl + Shift + . to cycle through the modes in this order: Ask → Run → Plan → PMO → Ask.
In a narrow window, the label collapses to an icon. Hover over it to see the current mode.
You can switch at any time,even while Agent is generating. The new mode takes effect on the next turn. Switching modesdoes not clear the conversation; Agent retains everything already discussed.
There is one side effect:switching modes cancels any pending approval request. If a card asking whether Agent may perform an action is waiting for you, switching modes dismisses it. Agent will ask again on the next turn if needed.
The four modes at a glance
| mode | Description in the app | In one line | Can it change things? |
|---|---|---|---|
| Ask | Search sources, analyze data, explain concepts, and run read-only queries without changing anything | Research and answer only | Never |
| Plan | Analyze requirements and design a solution without making changes | Draft a plan and wait for you to run it | Draft notes only |
| Run | Edit files, run commands, and complete tasks autonomously | Start working (default) | Yes |
| PMO | Break work down, track progress, and coordinate sub-Agents as a project manager | Manage the project and delegate to sub-Agents | Yes, through sub-Agents |
The composer prompt changes with the selected mode, giving you a second way to confirm it:
- Run — “Describe your task. XXX will get it done…”
- Ask — “Ask a question. I'll research and analyze it…”
- Plan — “Describe the goal you want to plan…”
- PMO — “Describe the goal. I'll break it down and coordinate sub-Agents…”
Run: the default mode for getting work done
New conversations start here. Agent can use every available tool: read and write files, run commands, edit docs and tables, call integrations, and delegate to sub-Agents.
Use it when: the goal is clear and the scope is controlled—for example, revising copy, adding a field, running a script, or implementing an agreed plan.
Remember: permission to work is not permission to do anything. Whether Agent asks before a risky action is controlled by the adjacentapproval leveldropdown (explained below).
Ask: research without changes—the safest mode
Agent becomes a read-only analyst. It can search, read files, inspect data, and run query commands such as database reads or log checks, butit cannot change anything. Writing or deleting files, changing business data, and running mutating commands are all blocked.
A few concrete boundaries:
- Itdoes not maintain a task list(it will not create todos and work through them); this mode is for answering questions
- It may delegate research to sub-Agentsin parallel, but those sub-Agents are also read-only
- Itcannot switch modes by itself. If your request is really “do this for me,” Agent will explain that the current mode is read-only and ask you to switch to Run or Plan yourself
Use it when: you only need to understand something—for example, “Why don't these figures match?”, “What does this logic do?”, or “How do competitors handle this?”
Plan: draft first, run only when you approve
This mode is for thinking before acting. Agent first investigates the current state with read-only tools, then produces astructured plancovering context and goals, current-state analysis, a recommended approach, risks and alternatives, and validation.
When the plan is ready, a card appears in the conversation titled Plan draft:
- Summary — a one-line overview
- Todos (N) — the task list, with each item marked Not started, In progress, Completed, or Cancelled
- Expand full plan — read the full plan
- Open document — read it as a document
- Run — the primary button in the upper right
Click Run, and two things happen:The conversation switches to Run automatically, and Agent begins implementing the plan. The button changes to Started(disabled), and a banner says “Started running ‘Plan name’.”
You always control execution. Agent cannot decide that a plan has been approved, start it on its own, or switch modes by itself.
What Agent may write in this mode: only .md notes and .canvas.tsx canvases used as drafts. It cannot write other file types or delete files.
To revise a plan: keep talking. Add criticism, constraints, or new requirements. Agent will draft a new card or update the current card's todos in place. Every historical plan card remains available and can be started independently.
PMO: split one goal into parallel workstreams
In PMO mode, the primary Agent stops doing all the work itself and acts as aproject manager: understand the goal → split it into two to five independent subtasks → assign a role to each → delegate → collect results → check them against acceptance criteria → summarize everything for you.
Tool access is the same as Run mode. The difference is that itdelegates concrete execution to sub-Agents by default. Independent subtasks runconcurrently.
After switching to PMO, a new button appears beside the composer:Team setup (choose collaborating sub-Agents). Open “Team for this conversation” and select the sub-Agent roles that may participate. With none selected, every role in the workspace is available. After saving, you will see “Team setup saved; takes effect with your next message.”
Once sub-Agents start, the interface showsActive sub-Agents. Open it to inspect each sub-Agent's execution stream.
Important boundaries:
- Each sub-Agent cannot see your conversation with the primary Agent. The primary Agent must include all context, inputs, deliverables, and acceptance criteria in the delegated task
- The approval level for a PMO conversationis fixed at Ask every time; it cannot be changed to Auto-approve or Allow all
- More sub-Agents consume more credits. If you see “Insufficient balance or this conversation's token quota has been exhausted,” reduce the number of sub-Agents
Use it when: a goal clearly contains several independent workstreams that will finish much faster in parallel.
Capability matrix
| What you need | Ask | Plan | Run | PMO |
|---|---|---|---|---|
| Research, read files, inspect data | Yes | Yes | Yes | Yes |
| Run commands | Queries only | Queries only | all | all |
| Edit files or write code | No | Only write .md / .canvas.tsx drafts |
Yes | Yes |
| Delete files | No | No | Yes | Yes |
| Create or change docs, tables, and other business data | No | No | Yes | Yes |
| Maintain a task list | No | Todos in the plan | Yes | Yes |
| Delegate to sub-Agents | Read-only sub-Agents only | Read-only sub-Agents only | Yes | Yes (the core workflow) |
| Produce a plan card with a Run button | No | Yes | — | — |
Best practices: choosing by situation
Before changing many files, align on a plan
Scenario: you want Agent to refactor a module, recalculate data under a new definition, or change a workflow that touches several areas.
What to do: switch to Plan, explain the goal clearly, read through the Todos on the plan card, and click Run only when the direction is right. Run.
Why: revising a written plan takes one sentence; undoing incorrect code or data is much more expensive. The plan card is also a shared contract between you and Agent, making misunderstandings obvious before implementation.
If you choose poorly: starting in Runmeans Agent reasons while editing. You may discover a faulty assumption only after it finishes, leaving a tangled set of changes that is harder to unwind than to redo.
When you only need an answer, use Ask—not Run
Scenario: “Why don't these two reports match?” or “What does this configuration do?”
What to do: switch to Ask and ask your question. It will investigate, build a table, and cite sources without changing a thing.
Why: this issafety enforced by structure, not a promise from Agent. The system blocks write operations in this mode, so Agent cannot change anything even if it tries.
If you choose poorly: in Run mode, asking “Why is this field empty?” may be understood as “fix it,” and Agent might fill it in when you only wanted the reason.
Use PMO when one goal can become several parallel workstreams
Scenario: “Review our competitors this quarter” requires research, data organization, and a written report.
What to do: switch to PMO, state the goal and acceptance criteria clearly. If needed, use Team setup to choose the participating roles.
Why: independent workstreams finish much faster when sub-Agents run concurrently, and each carries only the context it needs.
If you choose poorly: use Run forces the primary Agent to do everything sequentially. Context keeps growing, and earlier details start falling away.
For small, clear work, use Run and skip planning
Scenario: fix a typo, add a column, or run an existing script.
Why:Plan adds a plan document and requires another click to start. If the work itself takes two minutes, that extra planning step is unnecessary overhead.
Do not reach for PMO unnecessarily either: sub-Agents cannot see your conversation history, so the primary Agent must restate all context. For work that takes three to five steps, writing delegation briefs can cost more than doing the work.
Mode and approval level are different controls
The mode selectorTo its rightis another dropdown for theapproval level, which controls something different:
- mode controls what kind of work Agent may do
- approval level controls whether it asks before acting
There are three approval levels:
| Level | Description in the app |
|---|---|
| Request permission | Ask before actions |
| Auto-approve | Approve routine actions automatically; still ask for high-risk actions |
| Allow all | Run without asking for approval |
Do not guess which level is active. New workspaces default toAllow all. Trust the label shown beside the shield in the composer.
The two controls areindependent. You can use Run + Ask every time (Agent can work but asks at each boundary), or Run + Auto-approve (let it proceed on its own).
Two exceptions:
- PMO conversations are fixed to Ask every time; this cannot be changed
- Your organization may not allow relaxed approvals. If you see “Relaxed approvals are disabled for this organization; ask the owner to enable them in organization settings,” that is why
FAQ
Agent showed a card asking to switch modes. Should I accept?
This is expected. When Agent finds that the current mode cannot complete your request, it mayproposea switch, but itcannot change the mode itself. You must approve it.
The card is titled “Switch to X mode?” and explains what the change permits. Switching to Run, for example, says it will allow direct file edits and commands. The actions are “Switch to X” and “Cancel.”
Approve when it makes sense; otherwise cancel, and Agent must continue within the current mode.
If a Plan card is already waiting, the switch card notes that starting the plan will also enter Run mode. Start the Plan card directly instead of switching twice.
Will a new conversation use the mode I just selected?
Each conversation remembers its own mode. Switching conversation A to Plan does not change conversation B.
The system also remembersthe last mode you used with this Agentand uses it as the default for your next new conversation with that Agent. If a new conversation does not start in Run, you probably last used another mode with this Agent.
Why won't Agent make changes when I tell it to in Plan mode?
That is by design. Mode restrictions are enforced at thetool layer. Agent is not refusing; those tools are unavailable, and the system rejects any attempt to call them.
Saying “just change it; stop planning” cannot override this restriction. To let Agent act, switch to Runor click the Run button on the Plan card.
What happened to modes such as Learn and Yolo?
The current mode selectorcontains only these four: Ask, Run, Plan, and PMO.
Yolo is no longer a mode. How much freedom Agent receives is now controlled by the adjacentapproval levellevel. Learn mode is not exposed in the current app.
Does switching modes clear the conversation?
No. The entire conversation stays available and Agent remembers it. Only the tool permissions change, starting with the next turn.
The only thing dismissed isa pending approval request. Switching modes removes it, and Agent will ask again next turn if needed.
How can I tell which mode an older message used?
Your message bubbles show a small mode label in the lower-right corner for the mode active when each message was sent.
Only Ask and Plan receive this label—Run and PMO do not. Therefore, if a bubble hasno label, it was sent in Run or PMO.
What if a sub-Agent delegated by PMO gets stuck?
OpenActive sub-Agentsto inspect each sub-Agent's execution stream and see whether it is stuck or still working normally.
When a sub-Agent fails, the primary Agent analyzes the cause and may redelegate with a different role or strategy, degrade gracefully, or report the problem to you. “Insufficient balance or this conversation's token quota has been exhausted” means you should reduce the number of sub-Agents or check the account balance.
Why did Agent run a command in Ask mode?
Ask mode permitsread-only querycommands for inspecting data, logs, or status. They do not change anything.
Commands that create, change, or delete anything are rejected at the tool boundary and cannot run.
Still stuck?
Take a screenshot that includes the composer controls and send it to contact@larchiveai.com. We will help you diagnose it. Beta users can also ask their contact in the group.