Long Lake

For the organizational side of deployment, start with How do you diffuse AI into the real world?. Shenoy uses electricity adoption as an analogy for why better technology alone does not transform work. He describes Long Lake’s operator-owner approach and argues for embedding tools in existing systems, changing worker practices, and designing software alongside the people doing the work. This is the useful path for understanding the talk’s account of deployment responsibility and user enablement.

Return to Shenoy’s talk for its progression from copilots through synchronous, asynchronous, and long-running agents toward AI coworkers. He argues that autonomy must be earned through reliability and close iteration with users. Read this alongside his explicitly unresolved question: how can the parallel, sandboxed workflows of coding agents translate to traditionally serial services work? The recording offers a deployment framework and open problems, rather than evidence that every stage has been achieved.

For agent evaluation, use the same recording to trace Shenoy’s proposed improvement loop: capture real-work traces, evaluate against business outcomes such as closed books or repaired roofs, turn benchmarks into regression tests, and use operating data for customization and post-training. He connects this loop to enablement: usage supplies feedback, and improvement can encourage further usage. These are recorded proposals and examples, not independently verified performance results.