flâneur

Remember, Refine, Retrieve: A Context Engine for Enterprise Agents | Applied Compute

appliedcompute.com · 2,627 words · saved by 1 readers

Applied Compute builds Specific Intelligence for enterprises: AI systems trained on the institutional knowledge that makes their business unique. Specific Intelligence lives in two places: the weights of an LLM, and the context exposed to an agent at runtime. For the weights side, we run targeted RL over high-quality, integrated environments, among other techniques. We've written about this with DoorDash⌝, Cognition⌝, and Mercor⌝. Today we focus on the context side: how we create Contextbases that encode the nuance behind an enterprise's tasks, preferences, and procedures. We cover three topics: the architecture, benchmark results on enterprise tasks, and our most actionable finding — that context amortizes reasoning costs across agent runs. Most enterprise teams evaluating agents face the same tradeoff: higher reasoning effort produces better outputs, but drives up cost and increases latency. We show that using our Context Engine, we can shift where the tradeoff sits. An agent with ac

Introduction Applied Compute builds Specific Intelligence for enterprises: AI systems trained on the institutional knowledge that makes their business unique. Specific Intelligence lives in two places: the weights of an LLM, and the context exposed to an agent at runtime. For the weights side, we run targeted RL over high-quality, integrated environments, among other techniques. We've written about this with DoorDash ⌝ , Cognition ⌝ , and Mercor ⌝ . Today we focus on the context side: how we create Contextbases that encode the nuance behind an enterprise's tasks, preferences, and procedures. W

saved by

related reading