Watch a company actually make "the specification is the product, the implementation is generated" work, instead of stopping at the slogan.
The Prompt is the Platform - Dominik Tornow, Resonate HQ · Dominik Tornow
18 min total·Actually worth watching closely: ~5 min·3 must-watch clips
- 0:02 – 4:20Listen
The prompt is the platform: value moves from implementation to specification
Lays out the core claim: coding agents replace general-purpose platforms with bespoke implementations generated on demand, and reuse moves upstream to the specification; from 2026 the first software platforms will quietly retire.
What a company sells is no longer the general-purpose implementation but the specification, the protocol — and from that one protocol you derive implementations tailor-made for different infrastructure.
This stretch is spoken argument setting up the thesis; only the diagram at 3:45, on why the common mental model of agentic coding isn't enough, needs a glance.▶ Jump to 0:02Speaker · Dominik Tornow - 4:23 – 8:26Listen
Why direct generation fails: the agent has to move upstream
Explains the gap between abstract specification and concrete implementation: what the agent generates works on the happy path and passes basic tests, but breaks on concurrency, process failure, and network failure. Introduces the four-step pipeline: abstract specification, simulated implementation, concrete specification, concrete implementation.
Change the question: not "can the agent build the production system?" but "what does the agent need in order to design first and build second?" Humans stay in the loop, but the agent drives.
Pure conceptual argument with no visual dependency — fine to listen to on a commute; just hold on to the order of the four steps.▶ Jump to 4:23Speaker · Dominik Tornow - 8:30 – 12:08Skim
A real engineering case: stale reads on NATS
Takes Resonate's port of durable execution onto NATS: the protocol has been pared down to two objects, a durable promise and a durable task, yet the target's legal quirks — stale reads — still make a correct implementation hard.
The implementation must be correct when the target behaves legally, not merely when it behaves conveniently — a stale read is not a bug, it is permitted behavior.
Around 11:21 there is a slide explaining why only a failed write can reveal a stale read; worth pausing on the diagram to follow the causality. The rest is fine to just listen to.▶ Jump to 8:30Speaker · Dominik Tornow - 12:12 – 13:58Watch
Deterministic simulation: executable design
Shows the deterministic simulation testing environment built in Python: the simulated key-value store keeps a full version history per key, get returns an older version under control of the deterministic random generator, and update enforces optimistic concurrency.
The simulated implementation is not the product, it is executable design — it behaves like the real target in the ways that matter for correctness, but is additionally deterministic, repeatable, and inspectable.
From 12:32 the makeup of the simulation environment is shown; the design details on screen (version history, controlled stale reads) are the key to the method — worth watching screen by screen.▶ Jump to 12:12Speaker · Dominik Tornow - 13:58 – 16:14Watch
The forbidden fruit: giving the agent failures it can explain
The simulation records what the real platform hides: every get emits a trace event saying whether the read was fresh or stale, what you got, and what the latest value was. The algorithm may not depend on it, but the agent can use it to explain why the algorithm is wrong.
Once causality is visible, debugging information goes from "an invariant broke" to "it broke because the algorithm decided on a stale view of the world" — the agent doesn't just know it was wrong, it knows why.
The two trace-event comparison shots at 14:02 and 14:53 are the heart of the talk; a written retelling loses the contrast between hidden information and what production code actually sees.▶ Jump to 13:58Speaker · Dominik Tornow - 16:17 – 17:31Listen
Closing the loop: the agent becomes the design driver
Pulls the pipeline together: the agent builds a proof of concept in the deterministic simulator, derives the concrete specification from the algorithm already known to be correct, then derives the implementation. Minimalism is the finish line of three years of subtraction, not the starting point.
"The prompt is the platform, the specification is the product" does not remove humans from design; it lets the agent genuinely participate in design given an environment with deterministic feedback.
The close is a spoken summary with no new visuals; just listen through and connect it back to the opening claim.▶ Jump to 16:17Speaker · Dominik Tornow