Research and
retrospectives.
We publish when a decision is worth documenting.
These are not tutorials. These are not opinion pieces. These are records of specific engineering decisions we made in production AI systems, and what we learned afterwards.
-
The Glue Layer Became a Standard
Connecting a model to internal systems used to be bespoke glue — written once per integration, then rotting quietly. That layer is now a governed protocol, and the work moved rather than disappeared: which system an agent may touch, under whose authority, leaving what audit trail. What standardization solved for us, and the questions it handed back.
-
When the Frontier Converges
Within a few weeks of 2026, nearly every major lab shipped a new frontier model — and for the first time several of them land inside each other's error bars on most production work. If picking the strongest model is no longer a differentiator, the differentiator moves to what you build around it. What convergence changed in our selection and routing decisions, and what it did not.
-
After the Price Collapse
Inference now costs a fraction of what it did two years ago, and the reflex is to spend the savings on more tokens. We did exactly that, then went back and read our own traces. The binding constraint had moved — from cost per token to latency budget, context discipline, and the failure surface that more calls create. A note on what we re-architected once price stopped being the limit.
-
A Larger Window Is Not a Context Strategy
Context windows crossed a million tokens, and the first thing many teams did was stop pruning. Our traces say attention degrades long before the window fills, and that what you remove still matters more than what you can now afford to keep. Measurements from our own retrieval paths, and the pruning rules that survived them.
-
Design the Degradation Path First
Every AI feature has a bad day: the provider is slow, the retrieval index is stale, a tool returns something malformed and the model improvises around it. Most teams design the happy path and meet the other one during an incident. We write the degradation path first — what the system returns when it cannot be confident — and let that constrain the feature.
-
The Number That Does Not Survive Production
A benchmark reports a single-run success rate. Production asks the same agent the same question a thousand times, with tools that time out and inputs that drift. The gap between those two numbers is not noise — it is the part of the system nobody evaluated. How we score trajectories instead of outputs, and what that changed about which agents we shipped.
- Coming
Context Engineering Is the Product
Teams building on foundation models keep discovering the same thing, six months into production: the model was never the bottleneck. What enters the context window was. This note is what we learned when we stopped optimizing prompts and started architecting context.
- Coming
Evals Before Prompts
Most AI applications are built in the wrong order: prompt first, eval later (or never). The correct order is the reverse — define what "good" means in measurable terms first, then write the prompt that satisfies the evals. We explain why this order matters more than it seems.
- Coming
Model Routing as Architecture
"Which model should we use?" is almost always the wrong question. Production systems use multiple models routed by request characteristics. We discuss how we design routing, why benchmarks often mislead routing decisions, and how the routing layer itself becomes an application.