Knowledge Systems: The New GTM Stack — Jeffrey Wang, Exa
Exa’s internal agent clone combines writing examples, decision evals, and permission-scoped tools—a concrete pattern for encoding judgment without sharing full access.
Jeffrey Wang analyzed **760 emails** to model his writing style, including an **18-word average**, and derived evals from hundreds of prior decisions in Slack and email. The resulting Jeffbot can draft messages using company context.
Treat a personal or role agent as an evaluated system, not a persona prompt: derive behavior from real artifacts, calibrate against past decisions, and pair flexible chat with stable interfaces for recurring work. Keep purchased systems useful through agent-accessible APIs such as MCP.
Jeffrey Wang analyzed **760 emails** to model his writing style, including an **18-word average**, and derived evals from hundreds of prior decisions in Slack and email. The resulting Jeffbot can draft messages using company context. Treat a personal or role agent as an evaluated system, not a persona prompt: derive behavior from real artifacts, calibrate against past decisions, and pair flexible chat with stable interfaces for recurring work. Keep purchased systems useful through agent-accessible APIs such as MCP. Access must depend on the caller. Wang’s own use can read and write across more systems, while other employees receive a reduced tool set and draft-only behavior; the talk does not provide accuracy results for the decision imitation.
This turns role-agent design into an evidence and authorization problem: real communications supply behavioral evals, company systems supply context, and caller identity determines whether tools can read, write, or only draft. It confirms that durable memory and stable interfaces matter, while narrowing persona imitation claims because no decision-accuracy results are provided.