Designing reliable multi-step agents without losing your mind
A practitioner's taxonomy of failure modes and the control structures that contain them.
An evolving archive of technical research, design decisions, and operational lessons from building AI infrastructure.
Research is where loose questions get sharpened into operational hypotheses. Each area below is a long-running thread, not a topic page — open it to see what's currently being studied, what's been learned, and which build logs the thread feeds.
Start with the founding manifesto if you're new. Otherwise, follow any thread that maps to a problem you're actually facing. The archive is designed to be entered sideways.
Long-form investigations into the architecture, behavior, evaluation, and operation of production AI systems.
Reliable multi-step reasoning under real workloads.
Retrieval architectures for production knowledge work.
Graph control vs. handcrafted state machines.
Eval pipelines as an engineering discipline.
Durable state for agents and assistants.
Smallest defensible production envelope.
Observability for non-deterministic systems.
On fragmentation, the operator’s mindset, public iteration, and AI as leverage infrastructure — the thesis behind this lab.
A practitioner's taxonomy of failure modes and the control structures that contain them.
Hybrid retrieval, reranking, and the operational tradeoffs nobody publishes benchmarks for.
Treating evals like CI: traces, fixtures, regressions, and budgets.
Mapping cognitive science abstractions onto durable agent state.
Where general orchestrators win, and where they collapse under real workloads.
Trace shapes, sampling strategies, and what 'p99' even means for an LLM call graph.
The smallest defensible deployment surface for AI features.
Why most tool-using agents fail at the API boundary, not the model.