John and Garrett compare the workflows they've built around Nunebot and Workbench: deterministic checks, a second model looking for holes, a limit on review cycles, and pull requests that explain why a change matters and prove it works.

They also get into moving Fireside's file storage to Cloudflare R2 behind separate read and write flags, watching successful uploads instead of waiting for support emails, and throwing away an overgrown implementation to find a simpler one.

Along the way: personalized interfaces, the limits of read-only access to your email, AI-assisted appliance repair, and Garrett's basketball app. The point isn't to stop caring about software. It's to spend more time on the parts you actually like, with enough evidence to feel good about shipping.

Links:

• Don’t Make a Human Do What AI Is Amazing At — John’s write-up of Nunebot: small PRs, bounded review loops, and proof that changes work.
• Reviewer — Garrett’s tool for running automated code review checks.
• Flipper Cloud — Feature flags for gradual releases and safer deployments.
• Fireside — The podcast hosting platform behind the storage migration discussed in this episode.
• Honeybadger — Error monitoring used to find problems and watch deployments.
• Flare — Performance monitoring for Ruby and Rails applications.
• Episode 1: Changing the Wheels While the Car Is Moving — Rebuilding Fireside while keeping it running, verifying dangerous changes, and knowing what complexity is worth adding.

Two years later, almost every piece has been replaced while podcasts, feeds, and downloads kept running.

In the first episode of The Friday Deploy, they break down how they rebuilt Fireside while it was live, including a jump from Postgres 9.5 to 18 with no downtime for listeners. They talk about reverse-engineering undocumented infrastructure, using Flipper to put the application into read-only mode, testing migrations until they became boring, and letting agents carry work all the way to the final approval.

This is not a story about giving an AI agent one prompt and letting it loose on production. It is about breaking dangerous work into small, verifiable steps and knowing where experience and taste still matter.

They also get into:

• Why verification matters more than another round of prompting
• How multiple models can provide independent reviews
• Giving agents safe, read-only access to production context
• Keeping deterministic work out of nondeterministic agents
• Using real customer data to guide product and interface decisions
• Scheduled tasks as the beating heart of an AI-assisted workflow
• Whether the next generation of software needs a traditional interface at all

The goal is not to eliminate risk. It is to gather enough evidence that shipping stops feeling like a leap of faith.

Links:

• Fireside
• Flipper
• Flare