JevMade hello@JevMade.com
← Back to agent workflows videos

JevMade field notes / Video guide

Why I couldn't build Jev at OpenAI — Diogo Almeida, TypeSafe Co-founder & CEO

TypeSafe CEO Diogo Almeida discusses Jev’s machine-oriented API, the distinction between assistance and automation, and why he favors small, testable decisions over long conversational prompts. The interview also explores external memory and coordination in multi-agent coding systems.

Original by Latent SpaceAgent workflowsDeep dive142 min 22 sec Published Source reviewed

Before you press play

What you’ll find in the video

  1. Jev is designed as a machine-native System One model mapping outputs to programming control-flow primitives like enums, booleans, and sorting thresholds rather than conversational chat.
  2. Developers should decompose AI workflows into small, discrete, and measurable decision queries using structured JSON state instead of stuffing tasks into massive system prompts.
  3. Decoupling multi-agent coding systems from a single shared KV cache allows parallel agents to coordinate state through external software memory and structured subtask hierarchies.
Worth knowing

Auto-generated English captions reviewed with Gemini. Diogo notes Jev is fundamentally empirical, does not guarantee deterministic outputs across queries, and explicitly rejects public benchmarking scores in favor of private workflow evaluations.

Why I couldn't build Jev at OpenAI — Diogo Almeida, TypeSafe Co-founder & CEO