Delegate the work, keep the wheel. Agents pause for approval before anything risky, ask when only you can decide, and flag the blockers only you can clear, so nothing irreversible happens without your say-so.
An Actionist agent does the work, but you hold the wheel. It pauses for your approval before anything risky, asks when a decision is genuinely yours to make, and flags the blockers only you can clear. Delegate the outcome. Keep the final say.
Most of the time an agent just works. When it shouldn’t proceed alone, it does one of three things, each tuned so you’re interrupted only when it matters.
Approve
Before a risky or irreversible action (sending an email, charging a card, deleting a record), the agent pauses and waits for your go-ahead. You set how cautious it is, per agent.
Ask
When a choice is genuinely yours to make, the agent asks a short multiple-choice question mid-run and continues the moment you answer.
Flag
When something only you can fix stands in the way, like an expired connection or a missing key, the agent records a blocker and surfaces it under Needs attention with exactly what to do.
A scheduled run that hits a wall, and tells you exactly how to clear it.
An agent running on its own can’t invent a credential it doesn’t have. Instead of failing silently, it files a blocker with a plain-language fix and stops, ready to pick the work back up once you act.
actionist · scheduled run · ops
schedule fired · weekdays 09:00 · agent: Ops
opsPulling yesterday’s closed Linear tickets…
opsLinear returned 401 Unauthorized. The connection looks expired.
blocker raised · broken_connection · run stopped
opsWhat to do: Reconnect your Linear account in the Apps screen, then I’ll resume on the next run.
That same blocker now appears on your Home tab under Needs attention, with a one-click Fix button that takes you straight to the screen where you resolve it.
Every agent has an approval mode that controls when it pauses to ask you before using a tool. Set it once in the agent’s Tools tab. Skills and MCP tools respect it globally, and you can override individual tools below the setting. The four modes run from most hands-on to most autonomous.
Suggest
Ask each time
Guarded auto
Full auto
Propose tool calls; you run them manually. The agent plans and recommends each action but never executes on its own, so you stay fully hands-on. Best while you’re learning to trust a new agent.
Require approval before every tool execution. The agent does the thinking and queues each action for a yes/no. Maximum oversight with the agent still doing the heavy lifting.
Run safe tools automatically, ask for risky ones. Reads, searches, and drafts flow without interruption; writes and destructive actions pause for approval. The everyday default for most agents.
Run every tool without prompting. The agent acts end to end with no pauses. Reserve it for well-tested, low-risk routines you’ve already watched run.
Full auto removes every confirmation, including destructive actions. Use it only for agents and routines you’ve already verified by hand. For anything touching money, customers, or production data, prefer Guarded auto or Ask each time.
Open the agent's Tools tab
In Agent Studio, select the agent and open the Tools tab. Find Approval mode near the top. Its subtitle explains that it controls when the agent pauses before using a tool.
Pick a mode
Choose Suggest, Ask each time, Guarded auto, or Full auto. The setting applies to all of the agent’s tools, including Skills and MCP servers.
Override individual tools (optional)
Need one tool stricter or looser than the rest? Set a per-tool override in the tool list below the global picker.
Approval mode
Controls when the agent pauses to ask you before using a tool. Skills and MCP tools respect this globally; per-tool overrides live below.
When an action needs your sign-off, a request card appears with a plain summary of what’s about to happen and a preview of the change. Each request is graded by how risky it is, and the most dangerous ones make you type to confirm.
Safe write
A low-risk change, like a draft or a non-destructive update. Quick to approve at a glance.
Sensitive write
Touches something that matters, like sending a message or changing a record others rely on. Read before you approve.
Destructive
Deletes or overwrites. The card requires you to type to confirm before it will run. No accidental clicks.
Secrets are never shown in an approval card. Fields like tokens, passwords, API keys, and raw commands are masked, so you can review what an action does without exposing the credentials it uses.
Delete customer record
Destructive
Permanently delete Acme Corp (cus_8fK…) and all associated invoices.
When the call is yours, the agent asks, then keeps moving.
Some decisions aren’t risky, they’re just yours: which of two leads to prioritise, whether to use the formal or casual tone, which account to post from. Rather than guess or stall, the agent asks a short, structured question and continues the instant you answer.
Multiple choice, not an essay
Each question offers two to four clear options. Pick with a click or a number key. If free text might help, the agent includes an “Other” option. Some questions let you select more than one.
Asked at the right moment
The agent only asks to clear genuine ambiguity, confirm a specific choice, or pick between concrete options, not for open-ended chit-chat. A short status line tells you what it’s waiting on.
On an unattended run (a schedule or a trigger firing while you’re away), there’s no one to answer in the moment. Actionist tells the agent honestly that no user is available, so it makes a reasonable call or finishes and flags the decision, rather than hanging forever. For decisions that must be yours, route them through an approval mode or a workflow gate instead.
When an agent running on its own hits something only you can fix, it doesn’t just fail. It records a blocker: what’s stuck, why, and exactly what you need to do. Every open blocker shows up under Needs attention on your Home tab.
Each blocker is filed under one of six categories, so you can tell at a glance what kind of fix it needs:
Missing credential
A required credential or API key isn’t there yet. Add it in Credentials.
Broken connection
A connection is expired, revoked, or failing auth (for example, a 401). Reconnect the app.
Needs decision
A choice or approval only you can make. Open the agent and decide.
Missing config
Required setup or configuration is missing. Finish the setup on the app or agent.
Permission denied
The agent lacks the permission or scope to act. Grant access on the connected app.
Other
A user-actionable block that fits none of the above. The agent spells out the fix.
Takes you straight to the screen where you resolve it: the right app connection, or the agent itself. Actionist only ever links to a real in-app destination.
Blockers tied to an agent or a trigger clear on their own the next time that agent or trigger runs successfully. Once you’ve fixed the cause, the card disappears.
Cleared by you
Blockers rooted in a shared connection or app stay until you dismiss them or reconnect, because they affect every automation that depends on that connection.
Blockers are raised on unattended runs: schedules and triggers. In a live chat you’re already there, so the agent just asks you directly instead of filing a blocker.
Coming soon: Tasks for You. Needs attention shows blockers today, the things that stop a run. We’re extending it into a full inbox for non-blocking items too (“review this draft before I send”, “have a look when you can”, FYIs), so everything an agent wants from you lives in one place, sorted by urgency.
You don’t have to be staring at the dashboard. When an agent connected to a channel gets blocked, the same card (what’s stuck, why, and what to do) is delivered to your Telegram or Slack, with a link back to the exact fix screen.
Delivered to Telegram & Slack
A blocked run posts its What to do card to the agent’s connected channel, so a problem reaches you even when the app is closed and your laptop is shut.
One link back to the fix
Each card carries a link straight to the screen where you resolve it. No hunting through menus to find which agent needs you.
Coming soon: approve from the chat. Today the channel card notifies you and links back to the app. Next, you’ll be able to approve, reject, or answer right from Telegram or Slack with a single tap, clearing a blocker or signing off on an action without opening Actionist at all.
Can an agent do something destructive without asking me?
Only if you’ve set it to. Under Suggest, Ask each time, or Guarded auto, destructive and sensitive actions pause for your approval, and the most dangerous ones make you type to confirm. The only mode that skips confirmation entirely is Full auto, which you opt into per agent for routines you’ve already verified.
What happens to a scheduled run that gets blocked?
The run stops and files a blocker instead of failing silently. It appears under Needs attention (and in your channel, if connected) with the exact fix. Once you resolve the cause, an agent-scoped or trigger-scoped blocker clears on the next successful run; a shared-connection blocker clears when you reconnect or dismiss it.
Can an agent ask me a question while I'm asleep?
On an unattended run there’s no one to answer in the moment, so Actionist tells the agent honestly that no user is available. The agent then makes a reasonable call, or finishes and flags the decision, so it won’t hang. For decisions that must be yours, gate them with an approval mode or a workflow step so they wait for you rather than being decided automatically.
Do approval settings cover MCP servers and skills too?
Yes. The approval mode applies to all of an agent’s tools globally, including Skills and MCP tools. If you need one specific tool to be stricter or looser, set a per-tool override in the agent’s Tools tab.
Will I see the same blocker again and again?
No. Repeats of the same blocker are de-duplicated into a single card with a Seen N times count, so a recurring problem nudges you without flooding the list.
