Tornow warns code agents fail without deterministic simulation loops

Software development is approaching a structural tipping point. For decades, the industry relied on general-purpose libraries, frameworks, and platforms to abstract away complexity. Dominik Tornow, founder and CEO of Resonate, argues this paradigm is shifting. In a world where AI agents can generate code on demand, general-purpose implementations will increasingly give way to highly bespoke, minimal extensions of existing infrastructure.

This shift moves the core value of software engineering upstream. The product is no longer the concrete code implementation; it is the abstract specification. When the specification is the product, the prompt itself becomes the deployment platform.

The Failure of Direct Code Synthesis

When developers ask an AI agent to leap directly from an abstract specification to a production-grade concrete implementation, the results are routinely disastrous. During a joint initiative with Synadia to build a durable execution engine on top of NATS, the engineering team observed this exact limitation.

When tasked with building a Rust-based server on top of PostgreSQL, the AI agent easily solved the happy path. However, it completely failed to handle concurrency, network failure, or process termination. The generated system was a fragile prototype, not a production-grade system. Abstract specifications leave too many implementation decisions unguided, causing agents to make invalid assumptions about database behaviors and network consistency.

Shifting Agents Upstream via Simulation

To bridge this gap, engineers must introduce intermediate artifacts that allow the agent to participate in the design phase rather than just the coding phase. The breakthrough lies in a multi-step engineering pipeline:

Abstract Spec ──> Simulated Implementation ──> Concrete Spec ──> Concrete Implementation

Instead of requesting production code immediately, developers should task the agent with building a simulated implementation within a deterministic environment. This executable design serves to discover, test, and verify algorithms under partial failures and message reordering. Once verified in simulation, the agent derives a concrete specification, which finally serves as the blueprint for the production implementation.

Designing with the Forbidden Fruit

To succeed, agents require precise, immediate, and unambiguous feedback loops. In distributed environments like key-value stores, consistency anomalies like stale reads are legally valid under specific models but highly challenging to debug.

By leveraging a deterministic simulator built in Python, engineers can expose what Tornow calls the "forbidden fruit"—diagnostic data that real production environments hide. While production systems only receive a payload, the simulator records whether a read was fresh or stale, alongside the hidden latest value. This telemetry allows the agent to analyze exactly why a distributed invariant failed, empowering it to repair its own algorithmic errors before writing a single line of production code.