The following article contains recent clips Stay with Tim O’Reilly Interview. you can see full version On the O’Reilly Media Learning Platform.
eddie osmani He is one of my favorite people to talk about the state of software engineering with AI. He spent 14 years at Google leading Chrome’s developer experience team, and recently moved to Google Cloud AI to focus on Gemini and agent development. He is also the author of several books including O’Reilly effective software engineer (coming in March), and my co-host for O’Reilly’s AI CodeCon. Every time I talk to them I feel like I have a better handle on what’s real and what’s noise. Our recent conversation continues Stay with Tim O’Reilly There was no exception.
Here are some things we talked about.
The difficult problem is one of coordination, not of generation.
Eddy explained that there is a spectrum in how people are working with AI agents right now. On one side you have single founders running hundreds or thousands of agents, sometimes without even reviewing the code. On the other hand you have enterprise teams with quality gates, reliability requirements and long-term maintenance to think about.
Addy says that for most businesses, “the real limit is not necessarily having hundreds of agents for a task. It’s about organizing a modest group of agents that solve real problems while maintaining control and traceability.” He told that frameworks like Google agent development kit Now support both deterministic workflow agents and dynamic LLM agents in the same system, giving you the opportunity to choose when you need predictability and when you need flexibility.
The ecosystem is developing rapidly. A2A (Agent-to-Agent protocol Google contributed to the Linux Foundation) handles agent-to-agent communication while MCP handles agent-to-tool calls. Together they start to look like TCP/IP of the agent era. But Eddy was clear about where things stand: “Almost nobody has figured out how to make everything work together as smoothly as possible. We’re getting as close as possible. And that’s the real hard problem here. Not generation, but coordination.”
“Something Big Is Happening” Debate
We spent some time answering a question from the audience Matt Schumer’s viral essay Arguing that the current moment in AI is similar to that just before the COVID-19 pandemic hit. People who knew were sounding the alarm, but most people weren’t listening.
Eddy believed that “it felt a bit like being someone who wasn’t getting along, just finally getting around to trying out the latest models and equipment and having a memorable time.” He believes the piece lacked grounding in data and didn’t do a very good job of differentiating between what AI can do for prototypes and what it can do in production. As Eddy said, “Yes, models are getting better, harnesses are getting better, tools are getting better. I can do more with AI these days than I could a year ago. That’s all true. But to say that all kinds of technical tasks can now be done with near perfection, I personally don’t agree with that statement.”
I agree with Eddie, but I also know how it feels when you watch the future collapse and no one is paying attention. At O’Reilly, we started working with the Web when there were only 200 websites. In 1993, we created GNNThe first web portal, and the first advertisement for the web. In 1994, we conducted the first large-scale market research on the potential of advertising as the future business model of the Web. We kept lobbying for phone companies to embrace the Web and (a few years later) for bookstores to pay attention to the rise of Amazon, and no one listened. I’m a big believer in “something is happening” moments. But I am also well aware that it always takes more time than it appears.
Both things can be true. The direction and magnitude of this change are real. The models are getting better. Harnesses are getting better. But we still have to explore new types of businesses and new types of workflows. AI will not be a tsunami that will wipe out everything overnight.
Addy and I will co-host O’Reilly AI CodeCon: Software craftsmanship in the age of AI on March 26, where we’ll discuss orchestration, agent coordination, and the new skills needed for developers in more depth. We would love to see you there. Sign up for free here.
And if you’re interested in presenting at AI CodeCon, our CFP is open until this Friday, February 20th. See what we’re looking for and Submit your proposal here.
feeling productive vs being productive
There was a great line from a post named Will Manidis “Tool-shaped objects” Which I shared during our conversation: “The market for feeling productive is much bigger than the market for being productive.” The essay is about things that feel wonderful to make and use but that aren’t necessarily doing the job that needs to be done.
Addy noticed this immediately. “There’s a difference between feeling busy and being productive,” he said. “You can have 100 agents working in the background and you can feel like you’re being productive. And then someone asks, what did you make? How much money are you making from it?”
This is not to dismiss someone who is really productive at running a lot of agents. some people are. But it’s worth maintaining a healthy skepticism about your own productivity, especially when tools make it easy to feel like you’re moving faster.
planning is the new coding
Eddie talked about how the balance of his time on a job has changed drastically. “I can spend 30 to 40% of the time I spend on a job just writing down what I really want,” he said. What are the obstacles? What are the criteria of success? What is architecture? Which libraries and UI components should be used?
All the work you do to achieve clarity before you start code generation leads to higher-quality results from AI. As Eddy said, “LLMs are very good at keeping things down to the lowest common denominator. If there are patterns in the training data that are popular, they will use them unless you tell them otherwise.” If your team has established best practices, codify them in Markdown files or MCP tools so agents can use them.
I linked the planning phase to something bigger related to taste. Think about Steve Jobs. He was not a coder. He was adept at knowing what goodness looked like and inspiring those who worked with him to achieve it. In this new world, that skill matters a lot. You’re going to be like Jobs telling his engineers “no, no, not like that” and telling them what’s beautiful and powerful. Except now some of them are engineer agents. Therefore management skills, communication skills and taste are becoming core technical competencies.
Code review is getting harder
Eddy pointed out one thing that doesn’t get enough attention: “Growing teams are feeling like they’re being saddled with all this PR that’s generated from AI. People don’t necessarily understand everything that’s out there. And you get the increased velocity expectations from ‘What’s the quality bar?’ Will have to be balanced with. Because someone’s got to maintain it.”
Knowing your quality bar matters. What are the cases where you are comfortable merging AI-generated changes? Maybe it’s small and well divided and has solid test coverage. And what are those cases where you absolutely need in-depth human review? Clarifying that difference is one of the most practical things a team can do right now.
Yes, youth should still go into software
We received a question about whether students should still pursue software engineering. Eddy’s answer was emphatic: “If you’re someone who is comfortable learning, there’s never been a better time to get into software engineering. You don’t necessarily have to be burdened with decades of knowledge of how things have been built historically. You can look at it with very fresh eyes.” New entrants can approach the first agent. They can delve deeper into orchestration patterns and model trade-offs without forgetting old habits. And this is a real advantage when interviewing at companies that need people who already know how to work this way.
More importantly, in the early days of new technology people basically try to recreate old things. The really big opportunities come when we figure out what was previously impossible that we can now do. If AI is as powerful as it appears to be, there is no opportunity to make companies more efficient at the same old work. It aims to solve entirely new problems and create entirely new types of products.
I’m 71 years old and have been in this industry for 45 years, and this is the most excited I’ve ever been. Even more than the early web, even more than open source. The future is being reinvented, and the people who start using these devices are now becoming part of its invention.
token cost question
Eddy made a funny and honest confession: “There were weeks when I would look at my bill to see how much I was using in tokens and be just shocked. I don’t know if the productivity gains were really worthwhile.”
His advice: Experiment. Understand what your typical tasks cost with multiple agents. Extrapolation. Ask yourself if you would still find it valuable at that price. Some people spend hundreds or thousands per month on tokens and feel it is worthwhile because the alternative was to hire a contractor. Others are spending so much and feeling mostly busy. As Eddy said, “Don’t feel like you have to spend big bucks to not miss a win in productivity.”
I would add that we are in an era where these costs are largely subsidized. Model companies are covering the estimated cost to lock you in. Take advantage of it while it lasts. But also recognize that there is still a lot of work to be done on efficiency. Just as JavaScript frameworks replaced all hand-coding UI, we’ll find frameworks and tools that make agent workflows even more token-efficient than they are today.
The predictions for 2028 are already here
One of the most interesting things Eddy shared was that a group in the AI coding community that he is a part of made predictions for what software engineering will look like by 2028. “We recently revisited that list, and I was surprised to find that almost everything on that list is already possible today,” he said. “But how quickly the rest of the ecosystem adopts these things is on a much longer trajectory than is possible.”
The gap between capability and adoption is where most of the interesting work will happen over the next few years. Technology is outpacing our ability to absorb it. The real work right now is figuring out how to bridge that gap on your team, your company, and your own practice.
Agents are writing code for agents
Finally we answered another great question from the audience: will agents eventually produce source code that is optimized not for humans, but for other agents to read? Eddie said yes. Platform teams are already having conversations about whether to build an agent-first world where human readability becomes a secondary concern.
I have a historical affinity for this. I wrote the manual for the first C compiler on the Mac, and I worked closely with the developer who was hand-tuning the compiler output at the machine code level. That was about 30 years ago. We stopped doing this. And I am confident that there will come a similar moment with AI-generated code where humans will mostly let it go and trust the output. There will be special cases where people strive for perfect performance or correctness. But they will be rare.
That change won’t happen overnight. But the direction seems quite clear. You can help invent the future now, or spend time later trying to catch the people who do.
This conversation was part of my ongoing series of conversations with innovators, Stay with Tim O’Reilly. You can find previous episodes on Youtube.
