The piece argues that enterprises require systems functioning reliably under production conditions, not merely functional demonstrations. TinyFish proposes "codified learning" as a solution bridging the gap between probabilistic AI models and deterministic business outcomes enterprises demand.
Core Concept
Rather than treating workflows as monolithic tasks, the approach structures them as graphs of discrete decisions. Each node is typed, bounded, and measurable. As stated, "The system never 'solves the site.' It solves the next node."
This architecture enables parallel execution and result caching, allowing recovery from partial failures without replaying entire sessions. Cost structure shifts from scaling with browser time or token consumption to scaling with distinct decision quantities—many of which support memoization or inexpensive verification.
Node Composition
Nodes contain three elements: deterministic code handling transforms and validation; model-backed choices with protective guardrails; and codified heuristics—learned preferences documented, versioned, and safely replayed.
Business Benefits
The framework delivers reliability through narrow failure domains; throughput via node-level scheduling at fleet scale; predictable costs; and structured observability where traces map to business events rather than brittle scripts.
Illustrative Example
Checkout verification demonstrates the approach: breaking the process into product location, variant resolution, cart addition, shipping/promotion application, and total calculation isolates ambiguity to two nodes, keeping remaining operations deterministic and cacheable.