Agents that ask first
evolve runs scoped AI agents across the apps you already use. They do the reading, the lookup and the writing, and stop at anything that matters. You get a finished draft and one decision.
For founder-led software businesses with small teams. Onboarding a small group of design partners now.
you approved 0
Something happens
One agent, one lane
A finished draft
You decide
Done, on the record
Sample data and illustrative agents. If you don't press Approve, the demo moves on by itself. Edit and Reject work in the approval machine below.
What can an agent touch?
Each agent has a lane. The lane is enforced.
There is no do-anything agent. An agent is given a short list of what it may read and what it may propose, and the list is checked where evolve connects to your app, not inside the agent.
- A call outside the lane does not happen. The agent's good behaviour is not what stops it.
- Reading is free. Anything that changes something for a customer stops for you.
- The worst case is bounded by the lane, whatever the model does.
Press an action and watch where it stops.
Nothing sent yet.
Who decides?
You do, on a finished draft.
The agent that drafts never carries out its own draft. An approval sits between the two, and it is the same approval wherever you happen to be: the web inbox, a Slack message or the terminal. Your AI chat app is next.
- Approve it, edit it, or send it back with a reason. All three work here.
- You see why the agent wrote what it wrote, next to the draft.
- Every step lands in an append-only record, with a time and a name.
What is waiting for me?
One draft from support-agent needs your decision.
subject Checking in
- Usage down three weeks, two threads unanswered
- No open billing issue on the account
- Voice check passed, 0.92
Planned: review from an AI chat app such as Claude or ChatGPT, connected to evolve over MCP.
When does it stop asking?
When you say so, one kind of work at a time.
Trust is set per kind of task, on a four-step ladder. Everything starts at L0. After a run of approvals you did not need to change, evolve proposes the next step with the evidence, and you decide.
- Move the ladder and watch how much of the same work still reaches you.
- The system cannot promote itself, and you can move anything back down.
- Anything that moves money stays at L0 for good.
Which model does the work?
The smallest one that does it as well.
Each kind of work has a champion model and challengers. A challenger is replayed against your own past tasks, and when it matches you get the evidence and the choice. Frontier models keep the hard cases. Open-weight models, which you can host yourself, take the routine ones.
- Run the replay, then switch. Watch the cost of a task.
- A switch is per kind of work, and it can be undone.
- On a model you host, the draft never leaves your machine.
customer check-ins, 40 past tasks you approved
large model
championFrontier, hosted by a provider.
$0.024 a task
small model
challengerOpen-weight, yours to host.
$0.002 a task
Nothing replayed yet.
Figures are illustrative.
What you get.
Time back, because the work reaches you finished and your part is a click.
The reading across apps, the lookup and the writing are done before you see the draft.
Nothing goes out that you did not allow, so agents can work on real customers and real money.
Each agent is held to its lane, and anything consequential stops in your inbox.
It needs you less each month, on your terms.
Work you keep approving unchanged is put forward for more autonomy, and it moves up only when you say so.
It gets cheaper as it learns, because routine work moves to smaller models.
A model switches only after it has matched the large one on your own past tasks.
Same system, different job.
Six jobs a founder does by hand today. Each runs through the same five steps you watched at the top. Play any of them.
- support-agent
Quiet customer
Corvid Labs has gone quiet. Usage is down for three weeks.
A drifting customer hears from you the same morning.
- finance-agent
Failed payment
Birch Ltd's renewal failed twice. $290 is unpaid.
The reminder goes out the day it fails, not the day you notice.
- sales-agent
Overnight lead
Sara at Aster Co asked about SSO for 40 seats.
A lead that arrived while you slept has an answer and a meeting.
- finance-agent
Refund request
Maya's lamp arrived cracked. She sent a photo.
Maya has her money and an apology before Monday.
- founder-ops
Weekly digest
You asked what happened this week.
Your Friday review is written before you sit down.
- mender
Error alert
Checkout fails for carts with a zero-price item.
The fix is waiting, tested, when you wake up.
These are a few. It does more.
Examples of jobs evolve can be taught, not a feature list.
Finance
- Chase a late invoice
- Categorise the month's transactions
- Flag a contract about to auto-renew
Operations
- A supplier ships short
- Stock running low, reorder drafted
- Delivery delay notices
Marketing
- Reply to a public review
- Draft the launch post in your voice
- Welcome email for a new signup
People
- Onboarding checklist for a new hire
- Reply to applicants
- Turn meeting notes into follow-ups
The list has no end on purpose. The third time you walk evolve through a job, it proposes a workflow for it, so the catalogue grows from your own work.
How it differs.
- From a workflow tool
- You don't drag boxes to define automation. Agents file tasks and you approve them.
- From an agent framework
- A different layer. evolve runs scoped agents behind an approval gate and keeps the record.
- From a chatbot on your data
- It does work. Answering questions is the smaller part.
- From a helpdesk or CRM
- It has an inbox, and it is an approval inbox. The customer record stays in your app.
What keeps it safe.
Scope is enforced at the connection.
An agent's lane is checked where evolve connects to a tool, against a list of what that agent may call. A call outside the list does not happen.
Nothing consequential happens without an approval record.
The agent that drafts never executes its own draft, and actions on money or accounts stay at full approval for good.
Everything is logged and can be replayed.
Every draft, approval, edit, rejection and action is written to an append-only audit log with a timestamp.
A killswitch at three levels.
Halt one agent, one kind of task or one connector, from the web, Slack or the terminal. Lifting the halt needs a written reason.
Your apps keep their data.
evolve keeps the relationship layer: conversation history, approvals, notes and the audit log. The customer record stays where it is.
What leaves the machine.
When an agent calls a hosted model, the draft is sent to that provider. A model you host keeps it on your machine.
What is live today.
The machines on this page show how evolve is designed to work. This is what exists now. It is used by its founder, on one product, and it has not served an outside business yet.
Running today
- The approvals inbox on web, in Slack and in the terminal
- Approve, edit, reject, reassign and defer
- The append-only audit log
- The killswitch, per agent, per kind of task and per connector
- Escalation when a review sits too long
- Tools, skills and workflows the agent writes, published through an approval
In progress
- Trust-ladder promotion and demotion
- Champion and challenger model replays
- Further connectors, payments first
- Hardened containers per agent and an outbound firewall
- More than one reviewer
The evolve at month 12 is not the evolve at month 1. It's the one your business actually needed.
Request access.
We are onboarding a small group of design partners. Leave your email and, if you like, the first job you would hand over.
We use your email to reply about access and for nothing else. Privacy