Software Factory

Introduced the concept: building a real product with zero human-written code. Covered ground rules, the stack (Next.js, Supabase, Sentry, Vercel), harness engineering, and the SDLC stages agents will handle. Showed the empty repo and the Ona interface. No code written — that starts tomorrow.

Walked through the initial Next.js 16 scaffold with Supabase auth, Sentry monitoring, and a testing framework. Introduced the product: Memo, a Notion-style note-taking app. Showed the AGENTS.md approach and the PR Reviewer automation running live. Laid out the full automation roadmap. Page live at memo.software-factory.dev.

Ona CTO Chris Weichel joined as co-host. Walked through plan mode turning a basic brief into a detailed product spec, and the Feature Planner automation breaking that spec into sequential GitHub issues. By stream time, agents had autonomously closed 50 PRs and shipped the full initial feature set: auth, workspaces, pages, Lexical editor, full-text search, markdown import/export, and member invites. Introduced the two-loop automation pattern and progressive risk escalation. Empty shell to functional product in under a day.

Opened the automations dashboard and walked through the full stack: Feature Planner, Feature Builder, Bug Fixer, PR Reviewer, PR Shepherd, Verifiers, Incident Responder, and more — 14 automations chaining together from idea to shipped code. Showed a real PR where agents wrote, reviewed, and resolved comments without human input. Discussed what's still manual and what's fully autonomous. Lou introduced the 'in the loop to on the loop' framing and shared how Ona handles compliant auto-approvals at scale.

Ona's COO filed 12 bugs after stress-testing the app. The factory had been fixing bugs too — just not the same ones. Walked through the two worlds: Sentry catching runtime errors autonomously, while visual and interaction bugs need a human to point out. Showed the quality.md self-improvement loop, filed two bugs live on stream, and checked back at the end.

Ona's Field Marketing Lead joined as co-host and gave her honest first impression: "it looks like a Notion clone." Walked through the gap between AI-generated design and human-crafted design — AI reverts to generic patterns unless you push hard with rules and systems. Showed a Storybook design system built in the hour before the stream, moving from text-based specs to visual component guidance. Tried fixing contrast issues live with Claude Design — it improved the ratio but made everything look the same. Introduced the Week 2 plan: iterate on Storybook to give the factory taste, with humans directing design and agents implementing.

The factory ships faster than you can steer it. Explored the last manual step in the SDLC: operations back to planning. Discussed three layers of human input into a software factory: spec quality, feedback loops, and taste. Built a user feedback tool live during the stream to close the loop. Chris Weichel on the tension between speed and differentiation: "It's never been easier to feature bloat, and it's never been harder to differentiate."

Ona's Head of Product Matt Boyle joined to discuss how product management changes when the factory builds faster than you can steer. Reviewed a database feature built overnight from a 10-minute spec. Discussed the dopamine trap of prompting vs planning, why AI should raise the quality bar not lower it, and Ona's internal experimentation framework: ship fast, demo internally, two weeks to production or roll back. The question isn't what you can build. It's what you choose not to.

AWS Solutions Architect Shardul Vaidya joined from the London office alongside Lou to compare two real software factory architectures side by side. Shardul built his from scratch in 72 hours: a Rust state machine with adversarial agent review, dependency graph orchestration, and a nightly overseer that spots inefficiencies across all agent conversations. Compared approaches to planning, review, self-improvement, and governance. Lou drew the Kubernetes parallel: everyone's building their own factory right now, but standardization is coming.

The finale. Lou and Zach walked through the full project dashboard and SDLC map: 375 PRs merged, 50k+ lines of code, 16 automations, 1,067 tests, 87% autonomous rate, median PR-to-merge of 4.9 minutes. Showed the live app, traced the inflection points across 10 days, and reflected on what changed: "You don't build Memo, you build a factory to build Memo." Biggest learning: spec quality determines output quality, and the operations side of the factory needed more automations than the build side. The factory keeps running.