Agent Studio
Hire an agent in the next two minutes.
Describe a job, test it on a month of sample history, then decide how far you trust it.
Describe
Say what the job is, in plain words.
Draft
The studio drafts it for you.
Backtest
It replays a month of past decisions.
Correct
Each correction becomes a rule.
Promote
It earns its autonomy over time.
How it works
Hire it. Train it. Promote it.
Every agent is built from six parts you choose. Then it learns on the job and earns your trust.
Brain
The model that does the reasoning
- GPT-5.4
Job
Its instructions, in plain words
“Check every invoice against the PO. Flag any mismatch.”
Tools
The systems it can use
- ERP
- Slack
Knowledge
The documents it learns from
- Supplier terms
- Price list
- +12
Guardrails
Approvals, limits and undo
- Asks over €5k
- Task limits
- Undo 30 min
Trigger
When it starts work
- New email
- Schedule
- Webhook
Brain
The model that does the reasoning
- GPT-5.4
Job
Its instructions, in plain words
“Check every invoice against the PO. Flag any mismatch.”
Tools
The systems it can use
- ERP
- Slack
Knowledge
The documents it learns from
- Supplier terms
- Price list
- +12
Guardrails
Approvals, limits and undo
- Asks over €5k
- Task limits
- Undo 30 min
Trigger
When it starts work
- New email
- Schedule
- Webhook
Use cases
Hire for the work that eats your week.
Every team makes the same decisions hundreds of times a month. Those are the jobs an agent learns fastest, and earns trust in soonest.
Customer care
Refunds desk
Reads the ticket, checks the order and the photo, refunds within your limit and hands the rest to a person.
Decides
- Refund
- Ask for a photo
- Escalate
Acts alone, up to €50
Finance
Invoice checker
Matches each supplier invoice to its order and delivery, and holds duplicates and price changes before anyone pays.
Decides
- Approve for payment
- Hold
- Query supplier
Proposes, finance approves
Sales
Lead qualifier
Replies to new leads within a minute, asks what your best rep would ask, and books the call in the right calendar.
Decides
- Qualify
- Book a call
- Nurture
Acts and tells the rep
Operations
Dispatcher
Assigns each delivery to a driver by zone, capacity and shift, and replans the day when someone calls in sick.
Decides
- Assign driver
- Reorder stops
- Flag a late run
Acts and tells, 30 minutes to undo
Procurement
Freight buyer
Asks carriers’ agents for quotes and agrees terms inside the price, dates and volumes you set for it.
Decides
- Request quote
- Counter
- Commit
Commits inside its mandate
People
Onboarding coordinator
When an offer is signed, sets up the accounts, orders the laptop and plans week one with the manager.
Decides
- Create accounts
- Order a laptop
- Plan week one
Proposes, IT approves
Have a different job in mind?
Describe it in a sentence. The studio drafts the agent, the decisions it makes and the systems it needs.
Performance
Manage agents like staff, not software.
One page shows what every agent decided, how often your team agreed, what each decision cost and the hours it gave back.
PerformanceOctober · 6 agents
- Decisions
- 1,284
- +18% on September
- Agreed with your team
- 96%
- Target 95%
- Cost per decision
- €0.07
- €96 this month
- Hours given back
- 61 h
- At your minutes per decision
| Agent | Level | Decisions | Agreement | Cost each | Hours back | Trend |
|---|---|---|---|---|---|---|
| Refunds deskCustomer care | Act alone | 412 | 97% | €0.06 | 21 h | Rising |
| Lead qualifierSales | Act and tell | 318 | 95% | €0.09 | 16 h | Rising |
| DispatcherOperations | Act and tell | 276 | 98% | €0.05 | 14 h | Rising |
| Invoice checkerFinance | Propose | 189 | 93% | €0.11 | 8 h | Rising |
| Onboarding coordinatorPeople | Propose | 47 | 91% | €0.14 | 2 h | Rising |
| Freight buyerProcurement · new hire | Watch | 42 | 89% | €0.04 | None yet | Rising |
Cost per good outcome
Each decision is checked later in the system it changed. Was the refund disputed, the invoice paid, the call booked?
Agreement per decision
See which decisions it handles well, and which ones still need a person.
Demoted on its record
A run of declines or one bad outcome drops it a level, automatically.
Control
You stay in charge. Always.
Every safeguard runs in code, in front of every action, not in a prompt an agent could be talked out of.
Approve from where you already work.
When it needs a yes, the right person gets a message in Slack, Teams or email with Approve, Decline and Edit. If nobody answers, it moves to the next person.
Refunds deskAgent10:42
#refunds
Wants to refund €240 to order 4471.
Why: damaged on arrival, photo attached, customer since 2021. Over your €50 limit, so it is asking.
Also in
See the change first.
Every proposed action comes with a preview of exactly what it will change, field by field.
Order 44713 changes
- Status
PaidRefunded- Refund
- €240.00
- Note
- Damaged on arrival
Thirty minutes to undo.
When it acts and tells you, one tap puts things back the way they were.
Refunded €38, order 450229:41 left to undo
Limits it can’t argue with.
Amounts, counts and hours are checked in code before every action, not written in a prompt.
- Up to €50 a refund
- 20 refunds a day
- Weekdays, 08:00 to 18:00
€75 refund held for a person
A budget for every agent.
Set a monthly cap. It stops before it overspends and tells you why.
Refunds deskOctober
€84of €120
On track for €109 by the 31st
Every decision, on the record.
Who asked, what it saw, what it chose and why, who approved and how it turned out. Kept append-only, and yours to export.
Approve the refundOrder 4471 · €240
Good outcome
Proposed
10:42 · Refunds desk
Damaged on arrival
Approved
10:44 · Maya, in Slack
Over the €50 limit
Done
10:44 · Stripe
Refund re_3Pq…
Checked
17 Oct · 3 days later
No dispute opened
Append-only. A correction is a new entry, never an edit.
Agent Studio
Your next hire starts
with a sentence.
Describe the job. See its record on last month’s work. Decide how far you trust it.