Four cards, two operators, one converging observation: as AI handles more of the routine work, the difficulty and the bottlenecks do not disappear. They move. Simon Willison named what shifts for the individual engineer. Zach Lloyd named what shifts for the engineering organization.
What agents demand from the engineer
Simon Willison posted a short note on September 24, 2026, then anchored it with a keynote three days later. The two pieces fit together.
The September 24 note, captured in Coding agents shift the difficulty floor upward, leaving only the hard problems for human engineers, makes a counterintuitive claim about agent adoption: coding agents make the work harder, not easier. When an agent absorbs routine tasks, the tasks that remain are the ones the agent could not finish. The human engineer works on a filtered subset of the problem space, and that subset skews hard. The agent's output quality also introduces its own overhead: verifying correctness, catching subtle errors, and directing the agent well all require discipline that adds to, rather than subtracts from, the engineering effort on difficult problems.
"The more time I spend working with coding agents, the more convinced I am that they make software engineering even harder. We can do amazing things with them, but unlocking their full potential requires extraordinary discipline and knowledge."
Directing AI agents demands the same skill that defined good engineering. Defining goals clearly and choosing tools well is what engineering always was., from the September 27 keynote at WeAreDevelopers World Congress, explains why experienced engineers navigate the harder floor more effectively than junior ones. The skill that makes agent use effective is not new. Defining goals clearly, providing unambiguous instructions, and choosing the right tools: that combination has been the core of engineering since before AI existed. The model gives it a new name ("prompting") and a new surface, but it is the same underlying skill. Those who already had it find agents become multipliers. Those who lacked it find agents amplify confusion.
"defining goals, providing unambiguous instructions, and figuring out the right tools... is kind of what software engineering _is_."
Willison describes the overall effect with the Greg LeMond cycling frame: it does not get easier, you just get faster.
- Simon Willison, Coding agents shift the difficulty floor upward, leaving only the hard problems for human engineers. Agents filter easy tasks; human engineers inherit the hard residual.
- Simon Willison, Directing AI agents demands the same skill that defined good engineering. Defining goals clearly and choosing tools well is what engineering always was.. Directing agents demands clear goal definition and instruction precision, the same skill that separated good engineers before AI.
What software factories reveal about organizations
Zach Lloyd appeared on Lenny Rachitsky's "How I AI" podcast on September 21, 2026, describing how Warp ships 2,000 pull requests per month with their autonomous agent Wilson. The system takes a request from Slack, routes it through Linear for tracking, generates and tests a PR on GitHub, and measures human interactions per PR to identify where automation quality degrades.
A software factory is a connected pipeline from idea intake to merged PR, not a standalone code-generation agent names the structural claim: what Warp built is a cyclic pipeline from request intake to merged PR, not a code-generation agent. The pipeline is the system. Connecting the stages, tracking them, and using the data to improve each stage is what makes the throughput sustainable and compounding. An isolated agent generates a PR. The factory measures, learns, and improves.
The second card, When AI automates code generation at scale, human code review becomes the primary throughput constraint, identifies what the factory reveals once it runs. Once code generation is fast and automated, the bottleneck moves. It is now human code review. Warp's engineers do not spend time writing code; they spend time reviewing 2,000 PRs per month. That is where their attention concentrates, and that is the constraint that limits throughput. The implication for engineering leadership is direct: the next optimization is not more agents. It is review throughput.
Both Willison's individual-level observations and Lloyd's organizational-level observations point at the same structural fact: AI does not eliminate the hard work in software engineering. It moves it.
- Zach Lloyd, A software factory is a connected pipeline from idea intake to merged PR, not a standalone code-generation agent. A software factory is a full cyclic pipeline from intake to merged PR, not a code-generation step.
- Zach Lloyd, When AI automates code generation at scale, human code review becomes the primary throughput constraint. At 2,000 PRs per month, the bottleneck is not code generation but human review capacity.