The conversation opens with the OpenAI and Cursor drama, but quickly turns into the bigger strategic question: what happens when model providers can cut you off, prices keep dropping, and the gap between frontier and budget models keeps shrinking? The answer, according to the trio, is to become model-agnostic, provider-agnostic, and much more intentional about how you measure cost, quality, and task fit.

Chapters

• 00:00 — AI ecosystem updates and industry shifts
• 05:19 — How AI models reshape builder workflows
• 15:45 — Building internal AI agents and tools
• 23:47 — Building a better AI code reviewer
• 32:56 — Building adaptable software systems
• 35:27 — Giving agents the right tool access

Hosted by Prisma

x.com/prisma

prisma.io

Where to find us

Tyler Hogarth

x.com/tylerhogarth

linkedin.com/in/tylerhogarth

Søren Bramer Schmidt

x.com/sorenbs

linkedin.com/in/sorenbs/

Tyler Benfield

x.com/rtbenfield

linkedin.com/in/tylerbenfield

Takeaways

• Auto mode transition for Claude Code
• Use of auto mode for longer running agent processes
• Introduction of Muse Spark 1.2 and Muse Spark 1.2 contributor
• Implications of model routing and pricing for Meta's models AI and automation in software development raise concerns around compliance, standardization, and the role of personal and factory agents.
• The use of personal agents for idea refinement and the automation of the software development lifecycle can lead to improved productivity, quality, and experimentation.

I got back from the TypeScript AI conference and the thing that stayed with me was not a talk, it was the room. Non-technical founders trying to guide teams, designers moving into building, and an engineer who felt the industry had moved on without him, all in one place and none of them sure where to start.

Tyler Benfield, Platform Architect at Prisma, joins me to work through what happens next. We try to pin down what a software factory actually is, since nobody has written a spec for one, and end up comparing an assembly line to a handmade furniture shop. We spend a while on the Hugging Face security incident and the detail we think is being under-read: Hugging Face had to run GLM on its own hardware to defend itself, because the safeguards on the frontier models block cybersecurity requests. That leaves a question neither of us can answer cleanly. If a small team comes under the same kind of attack, what do they defend themselves with?

The conversation opens with a reframing. Most teams ask how to keep a human in the loop. The harder question is how to get the agent into the loop in the first place. The usual failure mode is access: you hand an agent documentation but no way to read live logs, inspect metrics, or act on what it finds. From there the two dig into the skills that carry over from coding agents to infrastructure agents. Treat every new agent thread as a new hire that needs real context. Keep docs indexed for fast retrieval during an incident. Use subagents to keep telemetry noise out of the primary agent's context.

With Tyler Benfield, who works on AI-native infrastructure at Prisma.

If you care about where infrastructure work heads as agents get more capable, this one earns your time. Subscribe for the follow-up.