Lessons from the AI Frontier

A talk on designing AI agent systems. The core thesis: curate ruthlessly — across context, tools, and output.

Curate your context

Context isn't about less or more — it's about the right context, carefully chosen. An average LLM context holds 200k tokens. If you expose 800 endpoints, you've burned a quarter of memory just saying hello. Think in checkpoints, not threads — summaries and decisions beat raw conversation history. And remember training data has a cutoff: tell it to web search.

Tools ≠ utils

Good engineering says: small, composable functions. That's bad tool design. Exposing getColumns(), setFilter(), setSortBy(), applyChanges() makes the agent into fragile glue across multi-step chains. Instead, design for outcomes: buildView(intent). One tool = one agent story. 5–15 tools, not 800.

Curate your output

AI output doesn't need to match what the product needs — build a translation layer. In the case study (a data grid assistant for nursery staff), asking the LLM to produce exact MUI column configs was unreliable. The fix: let AI return semantic categories (identifying, relevant, insight, hidden) and let deterministic code translate that into the exact shape. Pure functions, unit tested — mergeAIColumnVisibility() and anchorSystemColumns().

The pattern

Humans bring ideas, experience, and edge-case intuition. AI brings scale, tirelessness, and bulk implementation. The bridge between them is curated context, outcome-shaped tools, and a tested translation layer.