Triggers, agents, the tools and apps in use, and the outcome, as a flow graph. Moving lines are happening now. Dotted lines are handoffs.
Sending, replying, posting, publishing, booking, creating an event, deleting, submitting a form: all of it stops for a human yes. The rule lives in the system. Approved actions still fail closed on do-not-contact lists and spent budgets.
The Approvals page, inline in the chat, Telegram, WhatsApp, or a Slack channel. First answer wins. Approve, reject, request changes with a note, or drop single items from a batch. Runs pause cleanly and resume the moment you answer.
Switch abilities on or off per agent. Give an agent every app or a hand-picked list. Choose per app which actions are allowed and write instructions for how the agent should use it.
A daily limit on unattended spend. One switch pauses all unattended work while chat stays on. Low-balance alerts and alerts on sudden spikes.
Each agent's state at a glance: working, waiting on you, done, failed. A feed of steps as they happen and the upcoming schedule.
Every action every agent has taken, with whether it came from a chat, a schedule, a watcher, an incoming message, a goal or an approval, and what it cost.
Set your brand once: logo, colours, fonts, how you sound, a never-say list, boilerplate and email signature. Every agent, on every surface, follows it.
One living document about the business: what you do, products and prices, customers, competitors, channels, goals, the numbers that matter, people. Every agent reads it on every run. Agents fill it in as they learn and tell you when they change something. Every change is kept.
These are rules in the system, not requests to the model. A clever prompt cannot argue its way past them, and a yes from you cannot break the ones that protect your accounts.
Outward actions wait for a human. Enforced by the system, not the model. Money actions always ask, even for an agent you have told to work without asking.
Agents never type passwords, one-time codes or card details. They hand you the page. Keys are encrypted per workspace.
Text from a web page, a tool result or an incoming message is data, never instructions.
Do-not-contact lists, pacing floors and cool-downs apply underneath approval, so a yes cannot break them.
Every action is logged with its trigger and its cost.
A watcher or a heartbeat that finds nothing wakes no model and spends nothing.
A model on its own can only talk. A harness is everything that turns it into a worker, and it is the part most products skip. Otto is the harness.
Real tools, real files, real calls.
Working while you are out.
Work goes to the agent that owns it.
Anything that sends, replies, posts, publishes, books, deletes or submits a form stops and waits for your approval. That is enforced by the system, not by the agent's judgement. No agent can talk itself out of it.
Explore the control planeA live dashboard of every agent at work, an activity log of every action ever taken, and the cost of every run, one click away.
See the cost of every runNo card. Your keys or ours. Nothing sends without your yes.