Agenda

In practice, when Claude Code, Codex, and Cursor are all open at once, it usually means something dumber: copy-pasting the same bug report into three different chat windows, re-explaining context each time, and losing track of which agent already tried which fix. Is a meta-harness the solution? 

This talk digs into what that would actually take: combining multiple coding agents into a single workflow rather than three separate ones, setting spend caps and approval policies rather than manually babysitting every action, and sharing a live session so a teammate can review or co-drive rather than screen-sharing a terminal. 

Expect a live demo, a few lessons learned, and a verdict on whether it works.

Attendees will learn about GPU orchestration, advanced scheduling, autoscaling strategies and architectural decisions that minimize latency and maximize throughput. We’ll also cover storage considerations and practical approaches to deploying models in production. This talk is ideal for engineers and platform teams who want to understand how to operate AI workloads at scale while maintaining reliability and performance across cloud-native environments.

This session is a technical postmortem of a production agent that violated a non-negotiable business constraint without crashing, throwing an error, or missing context. The system appeared successful from the outside, but the outcome was wrong in exactly the way production teams cannot afford: it was helpful, confident, and unsafe.

We will examine why embedding business rules in prompts turns reliability into whack-a-mole. Regression testing becomes guesswork, model changes create unexpected behavior, and critical workflows depend on wording instead of enforceable constraints.

The session then walks through a practical pattern for deterministic control: move business logic back into the application layer, constrain execution with state-driven orchestration, wrap tool calls with validation layers, and test behavior by asserting state changes and tool usage instead of fragile text outputs.

Attendees will leave with a simple principle: let the model handle conversation, but let code enforce the rules.