Your next
hire is eager.
Train AI agents like new hires.
Hand it your files. Coach it in plain words. It writes its playbook, then gets to work in the software your team already uses.
$20 of free credit to try it. No card.
Built by teams from
See what your
coaching changes.
Choose a job. Run the example. Decide what the agent can do next.
From the files you give it to the approval it asks for, follow one request through its playbook.
“I was charged twice for order 58213. Can you help?”
“Check both charges. Anything over $50, ask us first.”
- 1Find the order and compare both charges
- 2Ask a teammate before refunding over $50
- 3Refund the duplicate and reply to the customer
- Teach
- Test
- Review
- Done
You know the business. It learns the ropes.
- 01
Bring the context.
Docs, spreadsheets, PDFs, old tickets. Folders and all.
- 02
Teach it your way.
Coach it in plain words. Read and edit the playbook it writes.
- 03
Test it first.
Try real requests. Check the answers and their sources.
- 04
Put it to work.
APIs, a browser or a computer. Inside the limits you set.
It writes its own playbook.
Your words become procedures, each tied to the lesson that taught it. Read every step. Change what needs changing. Nothing goes live until you publish it.
Choose a step to see how the conversation shaped it.
The coaching behind step 1
- 9:02YouDuplicate charges are our top ticket. Start from the order number.
- 9:02AgentI’ll find the order in billing from the number they give.
Your words become a procedure you can inspect and change.
Refund a duplicate charge
Goal: the duplicate is refunded and the customer knows when they’ll see it.
- 1Find the order in billing from the number the customer gives.
- 2Check both charges are for the same order and amount.If they differ, hand it to a teammate.
- 3Refunds over $50 wait for a teammate’s OK.
- 4Refund the duplicate, then tell the customer when they’ll see it.
You review it, edit it and test it on real requests before customers see it.
It gets cheaper as it learns.
Most requests a team handles are the same few, done the same way. Your evals prove which model each one really needs, and the steps that never change stop needing one at all. Big models only where they earn it.
Refund a duplicate charge
| Model | Passed | A conversation |
|---|---|---|
| GPT-6 Astra | 100% | 3.2¢ |
| Claude Sonnet 5.5 | 98% | 0.9¢ |
| GPT-6.1 SolPick | 97% | 0.6¢ |
| Claude Haiku 5.5 | 74% | 0.1¢ |
Use GPT-6.1 Sol for this procedure: within 5 points of the best, at 0.6¢ a conversation against 3.2¢.
81% less on this procedure. The rest of the agent keeps its model.
Refund a duplicate charge
- 1Find the order in billingNo AI
- 2Check both charges matchNo AI
- 3Over $50, wait for a teammate’s OKNo AI
- 4Refund it and say when it shows upNo AI
- Off the script? The model takes over from here.
- The model writes every reply
- 0.5¢ a conversation
- Steps without AI
- 0.06¢ a conversation
- The cheapest model that passes
- Run a procedure’s evals on a few models. Eager picks the cheapest one within a few points of the best, and one click moves that procedure onto it.
- Steps without AI
- Once a procedure works, turn it into fixed steps. Chat replies and browser clicks run as recorded, with no model writing them, and the model steps in only when a conversation leaves the script.
- Every cent on the record
- See what each conversation, each procedure and each model costs, in dollars and tokens, and where the steps ran without AI.
A little know-how. A lot of reach.
It gets what your team has: your files and playbook, the people it works with, your APIs, a browser, a computer and code. Bring your own computer, or let us host a VM for it (early access).
I was charged twice for my order.
Working on it…
Context it reads
Tools it uses
The work, ready for your team.
| Invoice | Amount | Result |
|---|---|---|
| INV-2046 | $1,280 | Matched |
| INV-2047 | $920 | Needs review |
| INV-2048 | $6,400 | Approval required |
Checked against purchase orders and AP policy.pdf
Eager to help. You’re still in charge.
Give it room to work, with the boundaries your security team will ask about. Built in, not bolted on.
Read about controls- Approvals
- Refunds, account changes and anything you mark wait for a teammate’s OK, in the conversation.
- Allowed apps and sites
- Each agent may use only the apps and websites you list.
- Limits
- Steps and time per task are capped, and anyone can stop a task at any time.
- Replay
- Every browser and computer step is recorded, so you can watch exactly what it did.
- Versions
- Changes land in a draft. You test them, publish them, and roll back if you need to.
- Audit trail
- Who changed an agent, ran a task or approved a step, person or API key, is on the record.
Start with a job you know inside out.
Pick the request your team handles most. Give it the context and teach it your way.
Sales copilot
Your reps ask. It knows every product, policy and portal, and says what to recommend.
Support agent
Answers your customers and fixes it in your systems, instead of linking an article.
Patient intake
Finds the chart, checks coverage in the payer’s portal, and tells the patient what they owe.
Freight
Reads the truck’s live location, converts its time zone, and moves the pickup in your shipping system.
Claims
Catches the swapped digit in a case number, reads the right report, and keys the claim.
Payables
Checks a bank-change request against the supplier’s number on file before anything changes.
Questions
What teams ask before they train their first agent.
What can it learn from?
Docs, spreadsheets, slides, PDFs and images. Drop in a whole folder when you train your first agent: no cleanup first.
What if our system has no API?
It uses the same screens your team does: websites in its own browser, and desktop apps on a computer. Use your APIs where you have them and the screen where you don’t.
Where do desktop apps run?
On your own Mac or PC, or a virtual machine in your cloud, with our desktop app. Or we host the virtual machine for you: hosted VMs are in early access, so ask for one.
Where does it answer?
On your website, in your own app or tools through the API, or behind the voice agent you already run, on Retell or any platform that can call an API.
How do procedures get made?
From your files and your conversations with it, and from its own first runs that worked. You can read and edit every step, and test it on real requests before customers see it.
How do we keep it safe?
You choose which apps and sites each agent may use. Sensitive actions wait for a teammate’s approval, steps and time per task are capped, and every step is recorded so you can replay it.
How do we keep costs down?
Two ways, both per procedure. Run its evals on a few models and use the cheapest one that still passes. And once it works, turn on steps without AI: the replies and browser clicks that never change run as fixed steps, and the model steps in only when a conversation leaves the script. You see the cost of every conversation and procedure as it happens.
What does it cost?
You pay for the tokens your agents use, from about 1¢ a task depending on the model, starting with $20 of free credit. See pricing, or talk to us about volume.
Good work starts with a little coaching.
Bring the knowledge. We’ll bring the eager new hire.