notes · patterns from building enterprise ai

07 published

Steal these patterns.

Things I learned shipping AI products that people use every day, written up as patterns you can adopt, not war stories.

01 · patternJune 2026

An AI agent isn't a product. The job to be done is.

Shipped a powerful 'ask anything' agent and watched it gather dust? The problem usually isn't the model. It's that you handed users a blank canvas instead of the answers they actually came for. Do it properly and your agents will thank you.

read the pattern

02 · patternJune 2026

Tool-first agents: using RAG to fix what NL-to-SQL can't

Frustrated by waiting 3-4 minutes for an NL-SQL agent that burns tokens and returns zero rows? The fix isn't a better prompt. It's retrieving valid filter values before the agent ever queries.

read the pattern

03 · patternJune 2026

You probably don't need a multi-agent orchestrator

Planner, router, five specialist sub-agents. It demos great. In production, every hop adds latency and failure surface. What moved adoption was the harness: tool quality, contracts, and boring reliability.

read the pattern

04 · patternJune 2026

Build the engine, not the deliverable

Shipping the same kind of thing over and over, each one needing to feel bespoke? Stop building instances. Build the engine that builds them, and make each one data.

read the pattern

05 · patternJuly 2026

Two clocks: running slow reasoning inside a fast conversation

Putting an LLM into a live voice or chat loop and watching per-turn reasoning add a second of dead air that kills the whole feel? The fix isn't a faster model. It's splitting the work onto two clocks and accepting that your agent reasons one turn behind.

read the pattern

06 · patternJuly 2026

Grounding as a leash, not a library

Everyone reaches for RAG to make the model know more. But in a white-labeled or regulated build, the danger is what the model already knows: the real client's name, a competitor's product, a capability you never built. Here retrieval's job is the opposite of teaching.

read the pattern

07 · patternJuly 2026

Close the validation loop, or the human work evaporates

Your AI cleans the data, a human fixes what it got wrong, you ship the file, everyone moves on. The corrections were the most valuable thing you produced and you just let them walk out the door. The win is not a nicer review screen; it is keeping raw, AI, and human-validated results in one place so every correction makes the next run smarter.

read the pattern