AI Builders Summit 2026

A year of shipping agent products taught us that context design outlives model choice. This talk covers the patterns that survived: compaction budgets, retrieval boundaries, and the observability you need before you can tune anything.

Generated code compiles, passes review, and still breaks production. We catalogued a year of agent-written incidents and built verification layers that catch what human review misses — without slowing anyone down.

Read replicas at the edge, single-writer semantics, and what happens when you stop pretending you need a cluster. Real latency numbers from three production migrations, including the one we rolled back.

Users do not want to watch JSON assemble itself. Patterns for progressive disclosure, optimistic rendering, and graceful repair when the model changes its mind halfway through a response.

Everything your team already knows about queues, retries, idempotency and backpressure applies directly to agent orchestration. A tour of production multi-agent failures and the boring distributed-systems fixes that solved each one.

Our monorepo CI hit forty minutes and engineers stopped trusting it. We rebuilt it around content-addressed caching and got to six. A decision framework for which incremental-build investments actually pay off, and which ones are cargo cult.

Non-determinism does not excuse jittery interfaces. Tokens, motion budgets and content design for products where the backend is a probability distribution and the user still needs to feel in control.

Provocative title, sober content: capability scoping, output mediation, and the three-layer defence we run in production. Includes live demos of attacks that still work against naive agents in 2026.

A fast tour of the ten decisions we would make differently, from partition keys to the migration window we scheduled during a holiday. Ten minutes, ten mistakes, no filler.

We built three internal platforms nobody used before one that stuck. The difference was never features — it was migration cost, escape hatches, and letting teams keep their own deploy buttons.

A short, practical walkthrough of how to read a failing evaluation trace: what to look at first, which signals lie, and how to turn one bad output into a reproducible test case.

Four practitioners on the gap between the demo and the deployment: the failures that only appear at scale, the support burden nobody budgets for, and what they would build differently. Audience questions from the start.