AI engineering is changing so quickly that even the people building the most advanced systems are constantly changing how they work.
In this episode of AI That Works, Dex Horthy (Founder, HumanLayer; coiner of "Context Engineering") and Vaibhav Gupta (Co-founder & CEO, Boundary; creator of BAML) recap the latest Unconference, where engineers compared approaches to software factories, coding harnesses, context engineering, evals, and autonomous coding. Despite wildly different workflows, one pattern kept emerging: everyone is still figuring it out, and the best teams are building feedback loops that let them change course quickly.
They dig into the growing divide between reading every line of AI-generated code and relying on automated systems to catch problems, and discuss deterministic linters, adversarial review, and skill-level evals along the way.
The conversation then moves into a real example of AI-assisted development, where Dex uses an agent to build a large feature overnight before breaking the resulting 20,000-line change into smaller, reviewable PRs. From there, they tackle where AI-generated plans, research, and specs should actually live when humans and agents need to work on them together.
TIMESTAMPS
Why AI Engineering Keeps Changing
What This Episode Covers
Meet the Hosts: Dex and Vaibhav
Can AI-Generated Content Actually Be Identified?
What the AI That Works Unconference Revealed
Why Unattended AI Coding Produces Slop
How Automated Feedback Can Reduce Code Review
Context Engineering, Coding Harnesses, and AI Infrastructure
Why Changing AI Reasoning Can Invalidate the Cache
Building Evals for Skills and Finding the Real Bottleneck in Software Factories
Why AI Feedback Loops Keep Getting More Sophisticated
How Design Patterns Can Make AI Code Worse
Where Should AI Teams Keep Their Plans?
Why AI Mistakes Compound as Systems Get More Complex
How to Make Adversarial Review Actually Useful
Building a Large Feature With an AI Agent
Why a 20,000-Line AI PR Still Isn't Ready to Merge
Turning an AI Prototype Into Reviewable PRs
Where Do AI-Generated Plans and Docs Actually Live?
Syncing Agent Files With a Shared Cloud Workspace
What a Shared Agent Directory Should Actually Do
The Hard Problem of Human Edits During Agent Work
Can AI Agents and Humans Edit Documents in Real Time?
Wrapping Up (And a Preview of Eval-Driven Dev)
RESOURCES
Dex Horthy — Founder, HumanLayer. linkedin.com/in/dexterihorthy
Vaibhav Gupta — Co-founder & CEO, Boundary. linkedin.com/in/vaigup
Avery Townsend — Engineer, Boundary. x.com/codeshaunted
Transcript
I've watched this one closely. Ask me anything about it and I'll point you to the exact moment.