Backed by Y Combinator

Cloned companies to train your agents

RL environments seeded from real company data.

We clone the apps
a company runs on

13+ business tools, cloned and loaded with a real company’s data. They’re connected, so one email in Gmail can update HubSpot, Mercury and Carta.

One goal requires nine months and seven apps

  1. 01The agent gets a goal“Close a $250,000 seed investment from Maya.”
  2. 02It works across the apps, like an employeeEach card below is one step, in the app where it happens. Arrows show what each step needs first.
  3. 03Code checks that every record agreesNot “did it reply politely” — did the signed terms, the bank and the cap table end up consistent? Here, one doesn’t.
  1. Gmail

    Email the investor

    “Can we meet about our seed round?”

  2. HubSpot · CRM

    Log the conversation

    Investor contact + email history

  3. Drive · Company records

    Read company metrics

    Revenue, runway and ownership

  4. Calendar

    Book the investor meeting

    Jan 14 · 10:30 AM · Maya + founder

  5. Gmail · Company counsel

    Confirm with the lawyer

    Counsel approves: $250,000 investment · $10m cap

  6. DocuSign · Agreements

    Sign the SAFE agreement

    Both parties sign: $250,000 investment · $10m cap

  7. Mercury · Banking

    Receive the investment

    $250,000 received · matches the SAFE

  8. Carta · Cap table

    Record the SAFE agreement

    Agent enters an $8m valuation cap

  9. Check the recordsMismatch

    Signed SAFE: $10m cap

    Carta record: $8m cap

    The wire is correct. The task fails until Carta matches the signed $10m cap.

Specifically for long-horizon tasks

Some tasks can’t be checked until they’re done. Closing March’s books means matching every bank line, and the receipts and payouts it needs arrive days later.

Task: close the books for March

412 tool calls across 7 apps, over 6 simulated days

The agent calls a tool The company does something: a reply, an upload, a payout settlingThe company replies, uploads or settles
  • Mercury
  • Stripe
  • Gusto
  • Slack
  • Gmail
  • Drive
  • Ledger
Tool-call trace across seven apps over six days. The agent's calls cluster on day 1, then on day 2 it messages employees and vendors; replies, uploads and Stripe settlements arrive over days 3 to 5, and the agent keeps matching until the books close on day 6.
Apr 1Apr 2Apr 3Apr 4Apr 5Apr 6

Pulls 1,140 bank lines and March payroll.

Asks 14 employees and 3 vendors for missing receipts.

Replies and uploads trickle in. Stripe payouts settle on Apr 4. The agent matches each one as it lands.

Books close. The checks run.

Checks, run once the books are closed

  • Every bank line is matched to a source1,140 of 1,140
  • Ledger cash ties to the Mercury balance$1,284,310.17 = $1,284,310.17
  • Payroll entries match Gusto2 runs · $412,880.00
  • Stripe payouts match recorded revenue211 of 212A $4,200 payout that settled on Apr 4 was booked twice, once as revenue and once as a refund reversal.

Real enterprises recreated for RL

Please contact us! We'd love to purchase your data or help you push the frontier of enterprise AI

A snow globe holding a miniature city block, drawn in blue dots

Get in touch

Connect with the Stitch founders

Schedule a meetingBook a 30-minute meeting