The Velocity Lab

Episode Summary Generic AI PR review agents from Copilot, Gemini, and Claude are a baseline, not a finish line. Dan and Dave break down the five things senior engineers actually look for in a PR — and why a custom review skill catches issues that the stock agents miss. Key Topics The "V zero" problem — why generic GitHub PR review agents only get you so far Five things senior engineers want surfaced first: what+why, views/APIs changed, business logic, breaking changes, and concerns from the standard review Stale PR titles and bodies — Claude rarely updates them across iterations unless you force it Breaking changes at the API and database layer: not minor outages, 100% outages How agentic coding goes sideways — random indexes, drive-by event systems, features Claude assumes you'll need Why posting inline review comments is the underrated capability most generic agents skip Notable Quotes "Having a generic PR review agent is not as valuable as one that's custom for your company." "It's not like a minor outage, it's a hundred percent outage if you mess that up." "Claude has a lot of agency. It'll just go add features that it assumes you'll need in the future." About The Velocity Lab Dave O'Dell and Dan McAulay work inside engineering organizations every day helping them ship faster with AI. No hype, no BS — just what's working in the field. Subscribe: RSS

Episode Summary AI transformation isn't a tooling problem. It's an organizational one. Dave and Dan dig into why bringing in AI without cutting bureaucracy, top-down processes, and the people who create them just gets you the same speed with new tools — and what to actually change so builders can build. Key Topics Why technical adoption alone won't make your engineering org faster How processes, architecture committees, and "trust theater" silently kill velocity Builders vs maintainers — when to prune, and why "palace guards" are a real role to retire The honest test: ask your top performers what it would take to 2x — then act on the answer Speeding up the upstream too: product, design, security, sales — same playbook Why ownership and sideways accountability beat top-down hierarchies for moving fast Notable Quotes "As an engineering manager, your main job is to get the hell out of their way and let the builders build." "You have palace guards who look good and protect everything. They don't really do much. And you have commanders who go out there and kick ass." "Move fast and build things — you can actually build quickly if you have the right system in place." About The Velocity Lab Dave O'Dell and Dan McAulay work inside engineering organizations every day helping them ship faster with AI. No hype, no BS — just what's working in the field. Subscribe: RSS

Episode Summary Dave and Dan walk through the questions a board or executive will ask about AI — from inventory and tooling to ROI, accountability, and competitive risk — and what answers actually hold up. Pro tip: if you can't answer them, your competitors probably can. Key Topics Do we have an inventory of where AI is operating in the company? — most orgs can't answer this The scattershot tooling trap — why one tool, used together, beats four licenses spread across the team Who's responsible when something goes wrong? — calculated risk + a no-fault, no-blame postmortem policy Measuring ROI — forget "10x feels"; the only number worth tracking is PRs deployed to production Are we keeping up with competitors? — a 20% efficiency gap is massive, and it compounds in months How do we know AI is doing what we think? — agents, gates, canaries, and tuning context surfacing Are we losing institutional knowledge? — the answer is to auto-update docs and runbooks every night Notable Quotes "I don't think board members care about culture very much. Bottom line — are you guys shipping?" "Pick one tool, make your whole company use the same tool. A bunch of individuals using AI is more productive, but a team using AI together gets you 10x." "If you're not investing and seeing your team adopt AI and ship faster, your competitors are. Maybe not now, but they will in three months." "Everything that Claude needs to know is the same thing that humans need to know." About The Velocity Lab Dave O'Dell and Dan McAulay work inside engineering organizations every day helping them ship faster with AI. No hype, no BS — just what's working in the field. Subscribe: RSS

Episode Summary Dave and Dan break down the upfront infrastructure work required before a single repo is ready for autonomous programming. They walk through the full prep checklist — monorepo structure, lazy-loaded docs, 90% test coverage with end-to-end tests, a dev environment that actually matches prod, canary deploys, and observability — and explain why each one is non-negotiable when you're shipping code without a human in the loop. Key Topics Why monorepos win — a root-level CLAUDE.md that knows how all your services interact, plus per-service docs underneath Progressive disclosure for context — keep CLAUDE.md under 200 lines and lazy-load topical docs (data model, GitHub Actions, Terraform, API endpoints) A daily docs cron — agents that update your docs every morning so context never goes stale Agent validation = tests + lints — 90% coverage as a hard gate, unit + end-to-end, all green before merge Shipwright crons in practice — development, review, docs, golden principles, cruft cleanup; one of them found and fixed an unauthorized Stripe webhook overnight Dev parity, canary deploys, and observability — the safety stack that makes shipping without a human safe (yes, just pay for Datadog) Notable Quotes "You can't just turn it on and expect magic to happen. There's a bunch of upfront work — we call it infrastructure — around your repository." "You're shipping code without a human involved. So you gotta put in as many safety things as possible. You gotta burn those extra calories to make sure dev is the same as prod." "I am an open source guy, but if it was up to me — just bite the bullet. Pay for Datadog and use APM." "By the way, these are things that would benefit any organization even if you aren't autonomous. Every org should be doing this anyway." About The Velocity Lab Dave O'Dell and Dan McAulay work inside engineering organizations every day helping them ship faster with AI. No hype, no BS — just what's working in the field. Subscribe: RSS

Episode Summary Dave and Dan share what they're seeing in the field this week: non-software engineers — in marketing, sales, customer service, and data — are getting massively accelerated by Claude Desktop and MCP servers. The episode centers on a real story of a single subject matter expert who built a Snowflake data tool, shared it across his entire company, and eliminated the need for one-off dashboard requests forever. The core argument: stop experimenting with every AI tool and pick one. For the whole company. Key Topics Claude Desktop for non-engineers — How marketing, sales, and ops teams are using Claude Desktop + MCP servers to do work that previously required engineering sprints The admin app problem — Why internal tools are always hacked together, and how giving Claude access to the data warehouse sidesteps that entirely Snowflake + Claude + Vector DB — How one team built a self-documenting data layer: Claude explores the warehouse, writes documentation, and surfaces the right tables via natural language Pick one tool — The case for committing your whole org to a single AI platform so shared tools, skills, and workflows actually compound across teams Building skills, not just prompts — Why the unlock isn't the first conversation with Claude, it's turning that conversation into a reusable skill that works every time What Anthropic hasn't built yet — Two missing pieces: scheduled tasks (a morning brief without cron hacks) and a direct bridge between Claude Code and Claude Desktop projects Notable Quotes "Be an adult. If you're in leadership, say: we are going with this tool. Dan and I recommend Claude Code." "It's not Claude, it's not GPT. These things are getting better and better. It's the system you're building on top of it." "A single human being who was an expert on this was able to create a tool for himself — and now he shared it with his team. Developers can use it. Marketing can use it. Anyone in the company can use it." About The Velocity Lab Dave O'Dell and Dan McAulay work inside engineering organizations every day helping them ship faster with AI. No hype, no BS — just what's working in the field. Subscribe: RSS

Episode Summary Dave and Dan announce Vitals OS — App Vitals' autonomous coding pipeline built on Claude Code and their Shipwright plugin. In the past two weeks it shipped 393 pull requests, with Dave and Dan writing fewer than 10 of them. They walk through how the pipeline works, where it still needs humans, and how they plan to bring it to client codebases. Key Topics The 393 PR milestone — what autonomous coding at scale actually looks like in practice The Shipwright plugin — a full DevOps-style pipeline covering research, planning, coding, and validation The 95% threshold — where the agent runs fully autonomous and where humans step in Slack as the interface — directing agents via voice notes while walking the dog The last 5% — deployment, monitoring, and architecture review as the human value-add Vitals OS for clients — launching the product for founders and non-technical builders Notable Quotes "Dan and I have done less than 10 pull requests out of the 400. It's pretty powerful stuff." "It doesn't stop until all the PR tests have passed successfully." About The Velocity Lab Dave O'Dell and Dan McAulay work inside engineering organizations every day helping them ship faster with AI. No hype, no BS — just what's working in the field. Subscribe: RSS

Episode Summary A client asked Dave and Dan a deceptively simple question: "What should we call the job title for the person who leads our AI engineering transformation?" The answer surprised them — it's not a new role. It's a Principal Engineer. In this episode, they break down exactly what that means, why the best candidate hasn't written a line of code in the last six months, and how to interview for it. Key Topics Why there's no new job title — agentic engineering isn't a sliver of your org, it's the whole thing. The role is Principal or Distinguished Engineer, full stop. The counterintuitive hiring criterion — you want someone who's written code for 10–20 years but hasn't written a line in the last six months. They've already built the trust with AI systems to let go. DHH and Linus flipped — both publicly opposed AI-generated code; now it's in Rails and the Linux kernel. The holdouts have adapted. So should you. Why rewrites almost always fail — Dave and Dan's take on migrations: the first 80% is easy, the last 20% kills the project, and you end up with something just as messy as what you started with. The interview process has changed — don't ask them to write code. Ask them to show you the system they built to build code. "If you built it by coding, you're not a good candidate." Management is in the way — autonomy and agency are what make senior engineers succeed. Micromanagement is the fastest way to lose the person you just hired. Notable Quotes "If you built it by coding, you're not a good candidate." "Your job nowadays is to not just build that system, but to constantly improve that system. You're building, maintaining, and improving a system that's building stuff for you — that is your job." "It's game over guys. You need to adapt if you haven't already." About The Velocity Lab Dave O'Dell and Dan McAulay work inside engineering organizations every day helping them ship faster with AI. No hype, no BS — just what's working in the field. Subscribe: RSS

Episode Summary Dave and Dan share a jaw-dropping collaboration breakthrough: Claude Code's voice mode accidentally picked up Dan's suggestions over a Google Meet call, turning a routine planning session into a real-time AI pair-planning session with two people and one AI. They also react to the OpenAI source code leak, discuss the real cost of their MAX subscription, and share how Claude Code fits into every part of their day — including 2 AM phone sessions from bed. Key Topics The Google Meet Epiphany — Claude Code voice mode picking up a second person's voice over video call unlocks true real-time collaborative AI planning Simpler than you think — Years of complex collaboration feature ideas, solved by speech-to-text + two people on a call The last 5% with clients — How this breakthrough eliminates friction between technical and non-technical collaborators OpenAI source code leak — Dave and Dan's hot take: it's more about the system you build on top of the LLM layer than the model itself The real cost of Claude MAX — $200/month subscription estimated at $5k/month in equivalent API usage Claude Code everywhere — Voice mode on walks, remote sessions at 2 AM, agents coding while you sleep Notable Quotes "You don't need a hundred different features. It's just get on a Google Meet, start pairing together." "Oh my God, dude, if we don't build some awesome product, it's our fault, not Claude." "If they were to actually charge what it actually costs, I might have to be an electrician." About The Velocity Lab Dave O'Dell and Dan McAulay work inside engineering organizations every day helping them ship faster with AI. No hype, no BS — just what's working in the field. Subscribe: RSS

Episode Summary Dave and Dan share how they shipped 254 pull requests in 7 days using AI agents Sully and Bodhi — with only about 10 hours of human effort combined. They unpack the key unlock (getting out of the terminal and into Slack), how they accidentally broke their own planning rules and recovered, and how adding metrics and learning loops turned a chaotic experiment into a real production system. Key Topics 254 PRs in 7 days: how autonomous agents maintained 40 PRs/day even on weekends Getting out of the terminal — why Slack became the unlock for async AI collaboration How two agents with the same model developed different personalities through memory Skipping the planning phase, chaos, microservices, and Wednesday night recovery Metrics as ground truth: moving from learning loops to hard numbers with PostHog What enterprise engineering orgs need to do right now (break your workflows) Notable Quotes "You should be breaking your workflows right now. You should be experimenting. You should not hold on to any legacy workflows." "We have two pull requests in the last 20 minutes, and Dan and I are sitting here talking. We did not do anything. It is just our bots out there working." About The Velocity Lab Dave O'Dell and Dan McAulay work inside engineering organizations every day helping them ship faster with AI. No hype, no BS — just what's working in the field. Subscribe: RSS

Programming languages have always built on top of each other. With agentic coding, is the knowledge of code itself going to disappear? Dave and Dan debate whether software engineers will need to know programming languages in five years, and what harness engineering means for the future.

Claude Code updates for March 27, 2026. Dave demos voice mode in the CLI, remote control sessions from your phone while walking the dog, and the 1 million token context window that somehow still is not enough.

Why giving every engineer access to every AI tool does not work. Dave and Dan break down the scattershot approach and why you need to focus on the entire SDLC.