A model for thinking about AI systems.
Just as software infrastructure decomposed into SaaS and PaaS, AI systems can decompose into layers. Each layer is independently buildable, deployable, and scalable.
Most AI systems are built as monoliths. They couple together retrieval, reasoning, orchestration, memory, and output.
When one breaks, everything breaks.
When you want to scale one layer, you scale all of them.
SAAL decouples this.
SAAL is not a prescription. It's a lens for diagnosing where systems break and why.
Start by naming which layer a problem lives in. The act of naming it precisely is half the diagnosis. A retrieval failure that gets misidentified as a reasoning failure gets fixed in the wrong place.
From there: define the contract between layers. What does the retrieval layer owe the reasoning layer? What format? What fallback? What latency budget?
Contracts make failures explicit. Explicit failures are fixable.
How the system accesses knowledge. Retrieval quality determines answer quality. This layer handles chunking, embedding, search, and context selection.
How the system routes, plans, and coordinates work. The control plane for agent behavior. Execution contracts, handoff protocols, failure handling.
Where inference happens. The model layer — but treated as infrastructure, not magic. Abstracted, swappable, observable.
How the system understands itself over time. Traces, evals, drift detection, re-grounding. The feedback mechanism for continuous improvement.
How the system interfaces with communities and creators. Community as infrastructure — not a marketing add-on but a product layer.
Building layer by layer, as constraints and problems demand solutions. Not all at once.
Most teams build retrieval as a tuning problem. It's a product surface. When it breaks in production, there's no visibility into why.
Multi-agent systems work until they don't. Without an explicit contract between agents, handoffs fail silently. Errors compound.
Audit isn't on the roadmap. Engineers want to ship features. Auditing last quarter's behavior feels like overhead. It is the work.
Community is bolted on. A Discord server added after launch. It doesn't integrate with the product and doesn't drive it.
A team applying SAAL's retrieval layer principles reduced document ingestion failures by isolating validation at document boundaries rather than chunk boundaries.
Defining explicit execution contracts between agents exposed silent failures at handoffs that had been misattributed to model quality for months.
SAAL is released under CC-BY 4.0. Use it in your decks, docs, team wiki — no permission needed. A link back is appreciated.