See it in Studio
What it looks like with this running.
Thursday · 4:12 PM
You4:12 PM
Every Monday, pull last week’s Stripe revenue, compare it to the quarter target, and put the summary in Documents.
Planner4:12 PM
Here’s what that compiles to.
Weekly revenue summary — 4 steps
Every step bound to a real tool: read Stripe, compare against the target on your Company page, draft the summary, save it to Documents. Nothing leaves your machine except the model calls.
Planner4:13 PM
One thing I couldn’t ground.
Which quarter target?
Your Company page has an annual target but no quarterly one. Add the real figure, or I can derive it as annual ÷ 4 and label it derived so nobody mistakes it for confirmed.
You4:15 PM
Rehearse it before you schedule anything.
It ran once against your live Stripe data while you watched. Output held for review — nothing saved, nothing scheduled.
Verifier4:16 PM
Clean run.
Ready to schedule
The rehearsal produced the summary in 38 seconds with no gaps and no failed steps. Mondays at 8 AM?
Scheduled. It shows up in Jobs with its own run history and trust bar.
End to end
The whole loop, start to finish.
Every step below runs today. Nothing here is a mock-up of what is coming.
- 1DescribeSay what needs to happen, in plain language. No form to fill in, no flow chart to draw, no code.
- 2CompilePencel turns it into concrete steps and binds each one to a real tool. Anything it cannot ground, it names instead of guessing.
- 3RehearseIt runs once against your own data while you watch, so you see real output before anything is scheduled.
- 4ScheduleGive it a cadence or a trigger. It becomes a job with its own run history and trust bar.
- 5WatchEvery run keeps what it read, which tools it called and what it produced, with a clean, repaired or failed verdict.
Capabilities
Everything it does, named.
One line each. No adjectives — every one is a specific thing you can go and watch it do.
- Takes a description of the work rather than a form, so you say what you want, read back what it built, and correct it in the same conversation.
- Compiles into named steps with a real tool bound to each, and reports any step it cannot ground in a tool you actually have instead of leaving the model to improvise.
- Runs once against your own data, with you watching, before a workflow may run unattended — nothing goes onto a schedule sight unseen.
- Gives every agent a role, a writing style, a preferred model, the skills it may use, behavioural restrictions and a risk tolerance, from asking before every action to acting alone and asking only when genuinely uncertain.
- Caps spending at both ends, with a maximum cost per run on an agent and a monthly credit cap on a room, neither of which can be exceeded.
- Stops at the approval gates you place, and stops at every external send and financial write whatever the workflow says.
- Catches up schedule fires it missed while the app was shut, refuses to overlap a run with itself, and can pause a job after repeated failures.
- Rolls recent runs into a per-job trust bar — clean, repaired or failed, plus the current clean streak — so you can see whether to leave it running.
- Presents a form at launch for any workflow you give inputs to, with nothing to design.
- Binds a step to Google Workspace, Slack, QuickBooks Online, GitHub, Notion, Atlassian, Linear, Stripe, a database, browser automation, or any MCP server you point it at.
Inputs and outputs
What it reads, and what you get back.
What it connects to
What lands in your hands
- A workflow you can actually read, compiled from a description
- A rehearsal run against your own data, before anything is scheduled
- A job with a schedule, a trust bar and full run history
- Agents you defined once, reusable across every Program
A week with it running
Day one to steady state.
What the first day looks like, what the quiet days look like, and when it comes to find you.
- Day 1Describe one thing you do every week. Watch it compile, then rehearse it once while you watch.
- When it looks rightPut it on a schedule. It becomes a job, with its own history and trust bar.
- Every runThe full trace is kept — what it read, which tools it called, what it produced, what it cost.
- Over timeThe trust bar tells you whether it has earned being left alone, instead of you guessing.
Guardrails
Where it stops for you.
The line between what it does on its own and what it will never do without you, stated plainly.
Waits for you
Anything it could not ground in a fact or a tool you actually have, and your go-ahead before a workflow moves from rehearsal onto a schedule.
Never does
Put a workflow on a schedule you did not approve, or bind a step to a tool you have not connected.