Benjamin Gruenbaum

Every team building agents hits the same wall: the interesting behavior only appears with real customer systems: Salesforce, Zendesk, Gong, Slack, GitHub, Snowflake. You can't develop, test, or benchmark against those. So we build mocks that pass and prove nothing and hit the wall in production.

We took the other route: 61 simulators of real vendor APIs, each also exposing an MCP server, all backed by seeded companies whose data is coherent across vendors and based on real production data - the account on a Salesforce opportunity is the same account a Zendesk ticket and a Gong call reference.

`docker compose up` and your agent has a "real" company to test with. We are releasing this for free.

I'll cover why building it cost over $200K in tokens, what was hard and what was easy. Why getting high quality data is incredibly challenging and briefly explain the science of the two papers we published about this.

Graph Engineering involves creating knowledge and task graphs. Creating a good knowledge graph involves multiple interesting steps like entity extraction, data parsing, noise elimination and graph algorithms. It turns out that if we start with a really good knowledge graph and do everything a company brain does in reverse we get a really good synthetic company with clones (twins) of all the common connectors we need. The missing piece was being able to tap into Eon's exabyte scale backup for metadata.