Summary
Dominik Tornow, founder and CEO of Resonate HQ, presents a thesis that goes beyond typical agentic coding discussions: coding agents will not just generate implementations — they will eventually replace the need for general-purpose platforms altogether by generating bespoke, infrastructure-native implementations from abstract specifications on demand. His talk uses Resonate’s own experience building a durable execution server on NATS.io (the open-source messaging system) as a concrete case study.
The key engineering lesson from that experience: when asked to jump directly from abstract protocol specification to concrete production implementation, agents consistently fail on edge cases. Resonate’s agent produced code that passed happy-path tests but broke under concurrency, process failure, and network partition. The fix required inserting an intermediate artifact — a “concrete specification” — where the human and agent collaboratively made all target-specific decisions explicit (database schema, indices, SQL queries, transaction boundaries) before any implementation code was written. Once those decisions were documented, the agent could reliably produce production-quality code.
The next evolution Tornow describes is moving agents even further upstream: giving them access to a deterministic simulation environment to discover correct distributed algorithms under partial failure before committing to a concrete specification. He argues this is only possible because Resonate spent three years shrinking their protocol to two core objects — a durable promise and a durable task — since complex protocols create state spaces too large for agents to reason about reliably. The talk closes with a reframing: the valuable product is no longer the implementation but the specification itself, with Resonate partnering with infrastructure providers including Senadia (Nats.io) to derive target-specific implementations from one shared protocol.
📺 Source: AI Engineer · Published June 29, 2026
🏷️ Format: Keynote Launch
