Max Headroom 
UNCLASSIFIED
On March 26, Addy Osmani and I are hosting the third O’Reilly AI Codecon, and this time we’re taking on the question of what software craftsmanship looks like when AI agents are writing much of the code.
The subtitle of this event, “Software Craftsmanship in the Age of AI,” was meant to be provocative. Craftsmanship implies care, intention, and deep skill. It implies a maker who touches the material. But we’re entering a world where some people with quite impressive output don’t touch the code. Steve Yegge, in our conversation earlier this week, put it bluntly: “Code is a liquid. You spray it through hoses. You don’t freaking look at it.” Wes McKinney, the creator of pandas and one of our speakers at this event, doesn’t write code by hand any more either. He’s burning north of 10 billion tokens a month across Claude, Codex, and Gemini, writing vast amounts of Go, a language he’s never coded in manually.
If that’s where this is headed, then what exactly are we crafting? That’s the question this lineup is built to answer, and the speakers come at it from very different angles.
The “dark factory” position
One end of the spectrum is occupied by people who are already operating what are increasingly being called dark factories, after the robot factories where there are no lights because the robots that do all of the work don’t need them. These are software production environments where humans set direction but agents do nearly all the implementation.
Ryan Carson is the clearest example on our stage. Ryan built and sold Treehouse, where he helped over a million people learn to code. Now he’s building Antfarm, an open source tool that lets you install an entire team of agents into OpenClaw with a single command. His talk, “How to Create a Team of Agents in OpenClaw and Ship Code with One Command,” is essentially a tutorial on running a software factory where a planning agent decomposes your feature request into user stories, each story gets implemented and tested in isolation by a separate agent, failures retry automatically, and you get back tested pull requests. This isn’t quite a dark factory, though. Ryan has built a CI pipeline where the agent records itself using a feature and attaches the video to the PR for human review. It’s an assembly line, and the human’s job is to inspect the output, not produce it.
This is Steve Yegge’s Level 7 or 8, and it’s no longer theoretical. But Ryan’s talk will also reveal what happens at the edges, when agents break, when the feedback loop fails, when automated retries aren’t enough.
The craftsmanship-means-oversight position
At the other end you have people who are deeply enthusiastic about AI coding but insist that the human role isn’t just “set direction and walk away.” It’s active, continuous, and skilled.
Addy Osmani anchors this position. His talk, “Orchestrating Coding Agents: Patterns for Coordinating Agents in Real-World Software Workflows,” is about the coordination problem. As he and I discussed in our recent conversation, there’s a spectrum from solo founders running hundreds of agents without reviewing the code to enterprise teams with quality gates and long-term maintenance to think about. Most real teams are somewhere in the middle, and they need patterns, not just tools. Addy has been thinking hard about what Andrej Karpathy called “context engineering,” the discipline of structuring all the information an LLM needs to perform reliably. His new book Beyond Vibe Coding is essentially a manual for this new discipline.
Cat Wu from Anthropic brings the platform maker’s perspective. She leads product for Claude Code and Cowork, and her focus on building AI systems that are “reliable, interpretable, and steerable” represents a design philosophy that the tool should make human oversight natural and easy. Where Ryan Carson’s approach pushes toward maximum agent autonomy, Cat’s work at Anthropic is about giving humans the right levers to stay meaningfully in the loop. I’m really looking forward to the conversation between Cat and Addy.
The costs of getting it wrong
Several speakers are focused squarely on what happens when the dark factory breaks down.
Nicole Koenigstein’s talk, “The Hidden Cost of Agentic Failure and the Next Phase of Agentic AI,” is about the failure modes that don’t show up in demos. Nicole is writing the O’Reilly book AI Agents: The Definitive Guide, and she’s been consulting with companies on the gap between what agents can do in a sandbox and what they do in production. Hila Fox from Qodo brings a complementary perspective with “From Prompt to Multi-Agent System: The Evolution of Our AI Product,” which traces the real path from a simple prompt-based tool to a production multi-agent system, including all the things that go wrong along the way.
The lightning talks share more results of real-world experience. Advait Patel, a site reliability engineer at Broadcom, will talk about what happens when AI agents break production systems, and how his team responded. Abhimanyu Anand from Elastic asks a question that should keep every AI builder up at night: “Is your eval lying to you?” If your evaluation framework is giving you false confidence, you’re building on sand.
The bottleneck was never hands on keyboards
Wes McKinney’s talk, “The Mythical Agent-Month,” revisits Fred Brooks’s famous argument that adding more people to a late software project makes it later, and asks whether the same dynamics apply to adding more agents. Wes’s answer, as he’s laid it out in his blog post, is so compelling that we immediately invited him to give it as a talk, even though that meant rearranging the existing program. Agents leave the essential complexity, the hard design decisions, the conceptual integrity of the system, completely untouched. Worse, agents introduce new accidental complexity at machine speed. Wes describes hitting a “brownfield barrier” around 100,000 lines of code where agents begin choking on the bloated codebases they themselves have generated.
This connects directly to something that Steve Yegge and Wes (and many others, including me) have converged on: Taste is the scarce resource. Brooks argued 50 years ago that design talent was the real bottleneck. Now that agents have removed the labor constraint, that argument is stronger than ever. The developers who thrive won’t be the ones who run the most parallel sessions. They’ll be the ones who can hold their project’s conceptual model in their head, who know what to build and what to leave out.
New architectures for the new reality
A cluster of talks addresses the structural question: If agents are doing most of the coding, what does the engineering organization, the platform, and the architecture need to look like?
Juliette van der Laarse’s talk, “The AI Flower: A Public Capability Architecture for AI-Native Engineering,” lays out a framework for how engineering teams should organize their capabilities in a world of AI-native workflows. Juliette’s work is a start on thinking through the second-order effects of the new technology. How does the organization itself need to change? We came across Juliette’s work recently and think it may be especially compelling for many of our enterprise customers.
Mike Amundsen has spent years thinking about API ecosystems and sustainable architecture, and he’s applying that lens to the question of how AI should relate to human expertise. His talk, “From Automation to Augmentation: Designing AI Coaches That Amplify Expertise,” makes a distinction that will determine the shape of the future human/AI economy. Automation replaces human work. Augmentation amplifies it.
Several other lightning talks fill in important pieces. Tatiana Botskina, a PhD candidate at Oxford and founder of an AI agent registry, talks about agent-to-agent collaboration and provenance, the question of how you know where an agent’s outputs came from. Neethu Elizabeth Simon from Arm addresses MCP server testing, a nuts-and-bolts reliability question that will matter more as MCP becomes the standard connective tissue for agent systems. And Arushee Garg from LinkedIn describes a production multi-agent system for generating outreach messages. These are all exploring issues that matter during real-world deployment.
The enterprise view
The event closes with my fireside chat with Aaron Levie, cofounder and CEO of Box. Aaron has been one of the most thoughtful enterprise CEOs on the question of what agents mean for SaaS and for knowledge work more broadly. His argument is that agents don’t replace enterprise software; they ride on top of it, and they need content, context, and governance to do anything useful. He’s also made the point that most companies have vast amounts of work they’ve never been able to afford to do, contracts they’ve never analyzed, processes they’ve never optimized. AI doesn’t just automate existing work. It unlocks work that was previously too expensive to attempt.
That connects to a theme I’ve been developing in my own work: the danger that AI creates enormous value but hollows out the economic circulatory system that supports the human expertise it depends on. Aaron is running a public company that has to navigate this in real time, making AI central to Box’s product while making the case that human judgment, context, and governance are more valuable, not less, in an agentic world.
What I’ll be watching for
There will be not only real excitement but hopefully deeper insight emerging from the tensions between these speakers and the positions they take. Ryan Carson and Cat Wu represent genuinely different philosophies of the human-agent relationship, and both are shipping real products. Wes McKinney and Addy Osmani agree that taste and design judgment matter more than ever, but they’re coming at it from very different vantage points: Wes as an individual developer pushing the limits of parallel agent sessions, Addy as someone thinking about patterns that work for teams of hundreds. Nicole Koenigstein and Hila Fox are asking the question that the enthusiasm sometimes papers over: What happens when it goes wrong?
And underneath all of it is the question that Steve Yegge, who isn’t on this program but whose ideas have certainly shaped my design of the program, would frame as a matter of grief and acceptance. Are we at the end of programming as a craft practice, or at the beginning of a new and different craft? I think the lineup proves that the craft isn’t dying. It’s migrating, from writing code to designing systems, from typing to taste, from individual heroics to orchestration. The people who understand that transition earliest will have an enormous advantage.
Sign up for free here. The event runs March 26, 8:00am to 12:00pm PDT.
AI & ML, Commentary
Radar
