Addy Osmani

A software factory is harnessing loops at scale - the system that builds your software instead of you. You can run the loop with humans in it (light factory) or you can ignore the humans (dark factory) and let those agents scope, build and ship code, without anyone really reading the details. But if people stop reading, they'll stop understanding your software. Your hardest job now is knowing which checks to build and how much autonomy to delegate.

Engineers need to own the outer loop, the accountability for these systems. A written version of my AI Engineer World's Fair 2026 closing keynote: quality, verdict, and answerability, the three hidden costs of delegation, and building software factories you can stand behind.

Taste used to be a byproduct of the reps. Agents took the reps. So if you're junior you now have to go get the taste and judgment on purpose. A builder's look at the weakening entry level, the two debts we now owe, and concrete ways to build judgment when the machine writes the code.

AI gets good at anything with an answer key. Your career is everything that doesn't have one.

The action in agentic engineering has moved from prompting to operating. Autonomy isn't one ladder, it's two axes (agency and orchestration) and six levels you move between per task. The real question is what level a task deserves, and what verification makes that level defensible.

I co-wrote a Google whitepaper about how AI is changing the software lifecycle. I'm not going to summarize the whole thing. Instead, here are the handful of ideas in it I think actually matter, plus six figures you're welcome to reuse.

Agents made writing code almost free. Understanding it costs what it always did, which is why review is now the bottleneck. The 2026 data is strikingly consistent on this, and yet most advice about AI code review is wrong for most people, because a solo developer with no users and a team maintaining a ten-year-old application are not solving the same problem.

You don't really need to be good at prompting anymore. The thing to get good at is the loop that does the prompting for you. It's five building blocks plus somewhere to keep notes, and Codex and Claude Code both have all five now.

Technical debt lives in your code. Cognitive debt lives in your head. Intent debt lives in the artifacts you never wrote - the goals, constraints and rationale that explain why the system is the way it is. It's the one kind of debt your agents can't pay down for you, and the one agentic engineering makes most expensive.

Starting more agents is easy now. However, more agents running doesn't mean more of you available - your cognitive bandwidth doesn't parallelize. All the judgement to actually steer them and merge the code they write into the codebase still has to route through exactly one serial processor which is just you. Orchestration tax is basically the price you pay for forgetting this and the only real fix is to start architecting your own attention like you architect any concurrent system.

Right now, it's too easy to let AI write the code while you skip the learning. The bug gets fixed. Your mental model doesn't move. We are silently trading future capability for present-day speed, and the tools won't force us to do otherwise. That part has to come from you.

Cognitive offloading is delegating to the AI and still owning the answer. Cognitive surrender is when the AI's output quietly becomes your output and there is nothing left to check. For software engineers the line between the two moves under your feet most days, and most of us are crossing it without noticing.

AI coding agents take the shortest path to done, which usually means skipping the specs, tests, and reviews that make software reliable at scale. Agent Skills encodes those senior-engineer behaviors as workflows the agent has to follow, with anti-rationalization built in.

A long-running agent can keep making progress over hours, days, or weeks. It can do this across many context windows and sandboxes, recover from failure, leave structured artifacts behind, and resume where it left off.

A coding agent is the model plus everything you build around it: prompts, tools, context policies, hooks, sandboxes, feedback loops. Harness engineering is the discipline of treating that scaffolding as a first-class artifact, and tightening it every time the agent slips.

AI coding agents consume documentation fundamentally differently from humans. If you're still optimizing only for human readers, you're leaving a growing share of your audience invisible to your tooling.

Running multiple agents in parallel is not just a question of throughput. It is a new kind of cognitive labor that requires managing multiple mental models, continuous judgment calls, and an ambient anxiety tax

The shift from conductor to orchestrator: how to coordinate teams of AI coding agents in real-world software workflows. From subagents to Agent Teams to purpose-built orchestration tools, this talk covers the patterns, tools, and discipline required to thrive in the era of agentic engineering.

Software engineering is not about writing code anymore. It is about building the factory that builds your software.

Two papers published in early 2026 suggest you might have just made your agent slower, more expensive, and no more accurate. The right mental model is to treat AGENTS.md as a living list of codebase smells you haven't fixed yet, not a permanent configuration.

Bias toward action is defaulting to the smallest responsible step that produces real feedback, while pre-committing to guardrails so that being wrong is survivable and quickly correctable.

More lessons learned from 14 years of engineering at Google, focusing on what truly matters beyond just writing great code.

Claude Code now supports agent teams - coordinated swarms of AI agents that research, debug, and build in parallel. What was feature-flagged is now real. Here's what it means and how to use it.