# The engineer of the future is the person who is able to choose what is worth doing. Addy Osmani | AI Engineer World's Fair 2026 | 18:26 Source: https://www.youtube.com/watch?v=n97BCfyFIvw Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/the-engineer-of-the-future-is-the-person-who-is-able-to-choose-what-is-worth Published: 2026-07-14 Tags: agents, coding-agents, engineering-culture, human-in-the-loop, workflows ## TL;DR - Engineers will be judged by the production decisions they can explain and own, rather than by how much code they personally write. - AI agents increase the need for evidence, verification, and human judgment because generated work can exceed the team's ability to review it. - The engineer's job is moving toward deciding what deserves attention, accepting risk, and taking responsibility for the system after it ships. ## Summary Addy Osmani argues that AI agents will automate more coding while making engineering judgment more important. Agents can investigate, implement, test, and report inside an execution loop. Humans still need to decide what is worth doing, inspect the evidence, approve or redirect changes, accept risk, and own the consequences. He describes the risks of cognitive debt, cognitive surrender, and orchestration tax. Teams can lose their understanding of a system, accept AI answers before forming their own judgment, or create more agent loops than they can manage. Osmani also frames careers in terms of alpha, the gap between a person's ability and current models, and decay, the speed at which that gap closes. Taste and judgment may last longer than narrow technical skills, but they also need to become visible through examples and decisions. His operating rule is simple: explain it or do not ship it. ## Key ideas ### Engineers will own the verdict on automated work [00:01](https://www.youtube.com/watch?v=n97BCfyFIvw&t=1s) Osmani defines the future engineer as someone who can choose what is worth doing and own the evidence, understanding, and verdict around agent-produced work. A verdict means being accountable for a production decision: deciding whether to ship, block, redirect, or accept a risk. Quality produces evidence, but answerability assigns responsibility for the decision. The question is shifting from a person's title to the part of the system they can own. Agents can help with modes such as prototyping, building, sweeping, growing, and maintaining, but people still choose which mode the product needs, what quality bar applies, and who owns the result. ### Harnesses and loops turn models into delegatable infrastructure [02:34](https://www.youtube.com/watch?v=n97BCfyFIvw&t=154s) The coding agent is more than a model. It includes the harness around the model, such as context, tools, the file system, and Git. Loop engineering adds repeated prompting, checking, remembering, and decisions about what happens next. At that point, agents start to feel like infrastructure. A software factory combines these parts, with agents working inside an inner loop and producing evidence. Humans remain responsible for production decisions. Osmani says the shift is toward making human judgment the highest-leverage checkpoint, because the system can generate work faster than people can understand and review it. ### AI-assisted code makes answerability an engineering requirement [03:06](https://www.youtube.com/watch?v=n97BCfyFIvw&t=186s) Osmani cites Sonar's 2026 survey to say that AI-assisted code is becoming a large part of many code bases. He also cites Sonar research finding that clean and messy repositories had roughly the same pass rates, while clean code used fewer tokens and caused fewer revisits. Maintainability therefore helps both people and future agents. The harder problem is review capacity. Osmani says 96% of people do not fully trust AI-generated code, while only about half always verify it before committing. Organizations then face questions about whether a model touched a file, which constraints guided the work, what evidence exists, what risk was accepted, and who owns the result. ### Taste is useful when it becomes visible judgment [06:15](https://www.youtube.com/watch?v=n97BCfyFIvw&t=375s) Osmani defines alpha as the gap between what a person can do today and what current models can do. Decay is the clock on that gap, because the model frontier can eventually reach a capability that once made someone distinctive. He agrees with Paul Graham that choosing what to make matters when anyone can make anything, but warns that taste can become a vague label for work people cannot yet explain. He uses Mitchell Hashimoto's definition of taste as making high-quality qualitative judgments when no objective metric exists yet. Taste matters when production becomes cheap, but it should produce critiques, examples, and decisions that teams and systems can learn from. ### Engineers must avoid cognitive debt and borrowed confidence [09:50](https://www.youtube.com/watch?v=n97BCfyFIvw&t=590s) Cognitive debt is the loss of understanding and memory about how to solve problems. In a code base, it is the gap between how much code exists and how much the team genuinely understands. A passing build and mergeable pull request can still leave a team unable to explain what it ships. Long-running agent tasks make this harder because a day-scale work stream cannot be reviewed with a glance at the end. Osmani also warns against cognitive surrender, where an AI answer becomes your answer before you form an opinion. He cites a Wharton study in which 73% of participants remained confident in a wrong answer after AI was wrong. ### More parallel agents create an orchestration tax [11:51](https://www.youtube.com/watch?v=n97BCfyFIvw&t=711s) Running more agents does not create more human cognitive bandwidth. Each additional loop creates decisions about routing, merging, verification, and integration. Osmani calls this orchestration tax. His proposed response is to design attention like a system: decide where a person enters the loop, what evidence they require, and what work or context they can reuse. The issue is not simply the number of agents. It is whether the person managing them can still make the decisions needed to understand and control the resulting system. ### High agency means choosing where to take ownership [14:13](https://www.youtube.com/watch?v=n97BCfyFIvw&t=853s) Osmani defines high agency as taking ownership of outcomes while knowing when to delegate, inspect, stop, and put your name on a result. It does not mean personally doing everything. His agency ladder moves from flagging a problem to executing, diagnosing, proposing, recommending, and resolving. At the top is discernment, deciding whether a problem deserves investment at all. Agents make more paths possible, so agency means choosing which paths deserve a person's ownership and attention. Agents can handle much of the inner execution loop, while humans remain responsible for the outer loop of deciding, verifying, approving, and owning. ### The shipping boundary is evidence and responsibility [16:13](https://www.youtube.com/watch?v=n97BCfyFIvw&t=973s) Osmani's operational rule is, "Explain it or don't ship it." Humans do not need to type every line or read every line, but someone must understand the work well enough to defend it. An agent can return diffs, tests, logs, rationale, traces, trajectories, or screenshots. Engineering begins when a person decides whether the work was worth doing, whether the evidence is sufficient, and whether to approve, redirect, or own the result in production. The boundary is therefore not simply a human looking at AI output. It is the point where evidence meets responsibility, much like an owners file identifies who is accountable for parts of a code base. ## Notable quotes - "I think that the engineer of the future is going to be really defined by the person who is able to choose what is worth doing." (00:43) - "If change is where humans enter the loop, if generation scales faster than comprehension, the scarce resource becomes judgment that's backed by evidence." (06:02) - "The failure mode is not using AI, but it's borrowed confidence." (11:25) - "Your cognitive bandwidth does not parallelize." (12:06) - "Explain it or don't ship it." (16:13) ## Tools & references mentioned - Boris Cherney - Sonar - Dex - Paul Graham - Mitchell Hashimoto - Wharton - Homer Simpson ## Who should watch - You are building agent harnesses or software factories and need to decide where human review belongs. - Your team is accepting AI-generated changes faster than it can understand, verify, or support them. - You are thinking about engineering careers and want a practical way to separate temporary model-resistant skills from long-lived accountability.