A whole company for your agent
Shaped like the 27 tools you already use
NetSuite
Plaid
Ramp
Pick a job. Watch it run.
You ask
In plain words.
We build it
About a minute.
It works the books
Over APIs it already knows.
It gets graded
Against a sealed answer key.
Follow the money.
One payment, every system it touches, to the penny.
{
"id": "ch_3Q8f2kLmN",
"object": "charge",
"amount": 482000,
"currency": "usd",
"customer": "Cobalt Systems",
"description": "INV-2287",
"status": "succeeded"
}It keeps moving.
Advance the clock. Invoices go out, payroll runs, cash moves, and the books stay tied out. Drag a dial and watch.
Ask anyone who works there.
They answer from the books, and only see what their job allows.
Hide anything. Keep the answer key.
Pick what's wrong. Your agent has to find it. Click one to try it.
Plugs into what you already use.
- Shaped like each vendor's own API.
- One header: your Fictix key.
- Signed webhooks, with retries.
- Save a company once. Everyone uses the same one.
Fresh environments.
Sealed answers.
Train and evaluate agents on real back-office work, graded by the data itself.
- A new company every seedSplit into training, eval and a private benchmark.
- Your agent can't peekThe answer key needs its own key scope.
- Graded by the dataPrecision and recall. No human labels, no model judge.
- Difficulty you setHow many problems, how well hidden.
- Task
- Find every bill that shouldn't be paid
- Tools
- QuickBooks · BILL · Plaid
- Grader
- Sealed answer key → precision / recall
- Seeds
- training · customer_eval · private_benchmark
- Episodes
- reset / reward API, on the roadmap
# works the books agent_key = ["read", "write"] # opens the sealed answer key grader_key = ["ground_truth"]
Better than a sandbox.
| Fictix | Vendor sandbox | Hand-made test data | |
|---|---|---|---|
| Full of data in | About a minute | Weeks of setup | Days |
| Realistic data inside | ✓A whole company | ✕Mostly empty | ✕A few rows |
| Books tie out across systems | ✓7 systems, to the penny | ✕One system each | ✕Rarely |
| Known answer key | ✓Sealed | ✕ | ✕ |
| Moves over time | ✓Advance the clock | ✕ | ✕ |
| Same data every run | ✓Saved by ID | ✕Drifts | ✓ |
Start free.
Every connector and the full API.
- 3 workspaces
- 500 test runs / month
- Webhooks + audit log
Pilots, contracts, admin.
- Unlimited workspaces
- Custom environments
- Flexible billing
Questions.
Is there any real company data inside?
No. Every company is generated from scratch, so there's no customer data to leak, scrub, or sign a DPA for.
Which systems does it look like?
27 of them, including QuickBooks, Xero, NetSuite, Stripe, Plaid, Ramp, BILL, Gusto, HubSpot, Slack and Gmail. The books tie out across QuickBooks, Stripe, Plaid, Ramp, BILL, Gusto and NetSuite.
Can I get the exact same company again?
Yes. Every company you keep is saved and gets a short ID like fix-7kq, so your whole team, and your test runs, use the same one.
How big can a company be?
Up to about 75,000 records for one legal entity. Very large companies are shown as a realistic sample, not a full copy.
Can labs use it for RL training?
Evaluation works today: sealed answer keys, precision and recall, and seeds split into training, eval and private benchmark sets. A reset and reward episode API is on the roadmap. Talk to us if you want early access.
Does my agent need special code?
No. Your agent calls the same APIs it already knows, with one host swapped, or reads CSVs. An MCP server is in early access.