On March 26, Addy Osmani and I are internet hosting the third O’Reilly AI Codecon, and this time we’re taking up the query of what software program craftsmanship seems like when AI brokers are writing a lot of the code.
The subtitle of this occasion, “Software program Craftsmanship within the Age of AI,” was meant to be provocative. Craftsmanship implies care, intention, and deep talent. It implies a maker who touches the fabric. However we’re coming into a world the place some folks with fairly spectacular output don’t contact the code. Steve Yegge, in our dialog earlier this week, put it bluntly: “Code is a liquid. You spray it via hoses. You don’t freaking take a look at it.” Wes McKinney, the creator of pandas and one in every of our audio system at this occasion, doesn’t write code by hand any extra both. He’s burning north of 10 billion tokens a month throughout Claude, Codex, and Gemini, writing huge quantities of Go, a language he’s by no means coded in manually.
If that’s the place that is headed, then what precisely are we crafting? That’s the query this lineup is constructed to reply, and the audio system come at it from very completely different angles.
The “darkish manufacturing facility” place
One finish of the spectrum is occupied by people who find themselves already working what are more and more being referred to as darkish factories, after the robotic factories the place there aren’t any lights as a result of the robots that do all the work don’t want them. These are software program manufacturing environments the place people set course however brokers do almost all of the implementation work.
Ryan Carson is the clearest instance on our stage. Ryan constructed and bought Treehouse, the place he helped over one million folks be taught to code. Now he’s constructing Antfarm, an open supply software that permits you to set up a whole group of brokers into OpenClaw with a single command. His speak, “The way to Create a Staff of Brokers in OpenClaw and Ship Code with One Command,” is basically a tutorial on operating a software program manufacturing facility the place a planning agent decomposes your characteristic request into consumer tales, every story will get carried out and examined in isolation by a separate agent, failures retry mechanically, and also you get again examined pull requests. This isn’t fairly a darkish manufacturing facility, although. Ryan has constructed a CI pipeline the place the agent information itself utilizing a characteristic and attaches the video to the PR for human evaluation. It’s an meeting line, and the human’s job is to examine the output, not produce it.
That is Steve Yegge’s Stage 7 or 8, and it’s now not theoretical. However Ryan’s speak will even reveal what occurs on the edges, when brokers break, when the suggestions loop fails, when automated retries aren’t sufficient.
The craftsmanship-means-oversight place
On the different finish you might have people who find themselves deeply smitten by AI coding however insist that the human position isn’t simply “set course and stroll away.” It’s lively, steady, and expert.
Addy Osmani anchors this place. His speak, “Orchestrating Coding Brokers: Patterns for Coordinating Brokers in Actual-World Software program Workflows,” is in regards to the coordination drawback. As he and I mentioned in our current dialog, there’s a spectrum from solo founders operating a whole bunch of brokers with out reviewing the code to enterprise groups with high quality gates and long-term upkeep to consider. Most actual groups are someplace within the center, and so they want patterns, not simply instruments. Addy has been considering onerous about what Andrej Karpathy referred to as “context engineering,” the self-discipline of structuring all the data an LLM must carry out reliably. His new e-book Past Vibe Coding is basically a handbook for this new self-discipline.
Cat Wu from Anthropic brings the platform maker’s perspective. She leads product for Claude Code and Cowork, and her concentrate on constructing AI programs which are “dependable, interpretable, and steerable” represents a design philosophy that the software ought to make human oversight pure and straightforward. The place Ryan Carson’s strategy pushes towards most agent autonomy, Cat’s work at Anthropic is about giving people the precise levers to remain meaningfully within the loop. I’m actually trying ahead to the dialog between Cat and Addy.
The prices of getting it improper
A number of audio system are centered squarely on what occurs when the darkish manufacturing facility breaks down.
Nicole Koenigstein’s speak, “The Hidden Value of Agentic Failure and the Subsequent Section of Agentic AI,” is in regards to the failure modes that don’t present up in demos. Nicole is writing the O’Reilly e-book AI Brokers: The Definitive Information, and she or he’s been consulting with corporations on the hole between what brokers can do in a sandbox and what they do in manufacturing. Hila Fox from Qodo brings a complementary perspective with “From Immediate to Multi-Agent System: The Evolution of Our AI Product,” which traces the actual path from a easy prompt-based software to a manufacturing multi-agent system, together with all of the issues that go improper alongside the best way.
The lightning talks share extra outcomes of real-world expertise. Advait Patel, a web site reliability engineer at Broadcom, will discuss what occurs when AI brokers break manufacturing programs, and the way his group responded. Abhimanyu Anand from Elastic asks a query that ought to hold each AI builder up at evening: “Is your eval mendacity to you?” In case your analysis framework is providing you with false confidence, you’re constructing on sand.
The bottleneck was by no means arms on keyboards
Wes McKinney’s speak, “The Legendary Agent-Month,” revisits Fred Brooks’s well-known argument that including extra folks to a late software program mission makes it later, and asks whether or not the identical dynamics apply to including extra brokers. Wes’s reply, as he’s laid it out in his weblog publish, is so compelling that we instantly invited him to offer it as a chat, regardless that that meant rearranging the prevailing program. Brokers depart the important complexity, the onerous design choices, the conceptual integrity of the system, utterly untouched. Worse, brokers introduce new unintentional complexity at machine pace. Wes describes hitting a “brownfield barrier” round 100,000 traces of code the place brokers start choking on the bloated codebases they themselves have generated.
This connects on to one thing that Steve Yegge and Wes (and plenty of others, together with me) have converged on: Style is the scarce useful resource. Brooks argued 50 years in the past that design expertise was the actual bottleneck. Now that brokers have eliminated the labor constraint, that argument is stronger than ever. The builders who thrive received’t be those who run probably the most parallel periods. They’ll be those who can maintain their mission’s conceptual mannequin of their head, who know what to construct and what to depart out.
New architectures for the brand new actuality
A cluster of talks addresses the structural query: If brokers are doing many of the coding, what does the engineering group, the platform, and the structure have to appear like?
Juliette van der Laarse’s speak, “The AI Flower: A Public Functionality Structure for AI-Native Engineering,” lays out a framework for the way engineering groups ought to set up their capabilities in a world of AI-native workflows. Juliette’s work is a begin on considering via the second-order results of the brand new know-how. How does the group itself want to vary? We got here throughout Juliette’s work lately and assume it might be particularly compelling for a lot of of our enterprise clients.
Mike Amundsen has spent years serious about API ecosystems and sustainable structure, and he’s making use of that lens to the query of how AI ought to relate to human experience. His speak, “From Automation to Augmentation: Designing AI Coaches That Amplify Experience,” makes a distinction that can decide the form of the long run human/AI financial system. Automation replaces human work. Augmentation amplifies it.
A number of different lightning talks fill in essential items. Tatiana Botskina, a PhD candidate at Oxford and founding father of an AI agent registry, talks about agent-to-agent collaboration and provenance, the query of how you realize the place an agent’s outputs got here from. Neethu Elizabeth Simon from Arm addresses MCP server testing, a nuts-and-bolts reliability query that can matter extra as MCP turns into the usual connective tissue for agent programs. And Arushee Garg from LinkedIn describes a manufacturing multi-agent system for producing outreach messages. These are all exploring points that matter throughout real-world deployment.
The enterprise view
The occasion closes with my hearth chat with Aaron Levie, cofounder and CEO of Field. Aaron has been one of the vital considerate enterprise CEOs on the query of what brokers imply for SaaS and for data work extra broadly. His argument is that brokers don’t substitute enterprise software program; they experience on high of it, and so they want content material, context, and governance to do something helpful. He’s additionally made the purpose that almost all corporations have huge quantities of labor they’ve by no means been in a position to afford to do, contracts they’ve by no means analyzed, processes they’ve by no means optimized. AI doesn’t simply automate present work. It unlocks work that was beforehand too costly to try.
That connects to a theme I’ve been creating in my very own work: the hazard that AI creates monumental worth however hollows out the financial circulatory system that helps the human experience it relies on. Aaron is operating a public firm that has to navigate this in actual time, making AI central to Field’s product whereas making the case that human judgment, context, and governance are extra useful, not much less, in an agentic world.
What I’ll be awaiting
There will probably be not solely actual pleasure however hopefully deeper perception rising from the tensions between these audio system and the positions they take. Ryan Carson and Cat Wu characterize genuinely completely different philosophies of the human-agent relationship, and each are transport actual merchandise. Wes McKinney and Addy Osmani agree that style and design judgment matter greater than ever, however they’re coming at it from very completely different vantage factors: Wes as a person developer pushing the bounds of parallel agent periods, Addy as somebody serious about patterns that work for groups of a whole bunch. Nicole Koenigstein and Hila Fox are asking the query that the keenness generally papers over: What occurs when it goes improper?
And beneath all of it’s the query that Steve Yegge, who isn’t on this program however whose concepts have actually formed my design of this system, would body as a matter of grief and acceptance. Are we on the finish of programming as a craft follow, or firstly of a brand new and completely different craft? I believe the lineup proves that the craft isn’t dying. It’s migrating, from writing code to designing programs, from typing to style, from particular person heroics to orchestration. The individuals who perceive that transition earliest could have an infinite benefit.
Join free right here. The occasion runs March 26, 8:00am to 12:00pm PDT.
