Writing the code has stopped being the slow part. Deciding what to build and checking what got built are the slow parts now, and both start with the plan.
A well thought-out plan is worth more today than at any point in my career, because building the thing has stopped being the slow part.
Faros AI, an engineering analytics vendor, compared periods of low and high AI adoption across 22,000 developers. Task throughput per developer went up 33.7 percent. Median time in code review went up 441.5 percent. Deployments per week went down 11.7 percent. In that data, more gets written and less gets shipped.
The labs that build the models describe the same shift from the inside. Anthropic’s economists call human attention “a constant bottleneck” in their own company and report that about 80 percent of interactive agent session time is spent idle, waiting for a person. The Pragmatic Engineer’s July 2026 report on Anthropic has engineers saying that implementation took about 15 percent of the time in the Bun rewrite. Sundar Pichai said in April 2026 that 75 percent of new code at Google is AI-generated and approved by engineers.
The constraint has moved to the two ends of the work: deciding what to build, and checking what got built. Sprint planning was designed for the middle.
Why short planning horizons made sense
Planning in small steps was a sensible answer to expensive execution. Barry Boehm’s research suggested that a problem found after delivery could cost up to a hundred times more to fix than one found during requirements. Boehm later qualified this himself, to “often” a hundred times and closer to five to one for small, non-critical systems, and a 2016 study of 171 projects found no evidence of the effect. The belief still shaped a generation of process design.
If late changes cost a fortune, you decide as much as possible up front. Kent Beck called waterfall “a rational reaction” to that cost curve, in the same 1999 article where he bet that the curve could be flattened. If change stays cheap, you can decide late, in small pieces, and learn as you go. That bet fed into the Agile Manifesto of 2001, with its preference for “responding to change over following a plan,” and into Scrum, which limits a sprint to one month or less on the grounds that longer horizons add complexity and risk.
Both sides of that debate assumed that building was slow. Two weeks of backlog was about as far as a team could usefully plan, given how long it took to build anything.
What changed
Agents can now carry a much longer piece of work on their own. The research nonprofit METR measures the length of task that AI agents complete without help, counted in the time a human expert would need. For GPT-4 in March 2023 that was about four minutes. For Claude 3.7 Sonnet in February 2025 it was an hour. For Claude Opus 4.6 in February 2026 it was about twelve hours. METR estimates a doubling time of roughly four months since 2023.
Vendors report longer runs in practice. Cursor described agents working 25 to 36 hours on a single task in February 2026. When a single run can cover what used to be days of human work, a plan that only reaches the next few tickets is too small a unit.
How the people furthest along plan
The companies with the most agent experience have moved toward more planning before execution.
Cursor’s long-running agents, by the company’s own account, propose a plan and wait for approval before they start, because a slightly wrong assumption at the beginning becomes a wrong result a day later. OpenAI published a 25-hour Codex run in February 2026 that was steered by four markdown files: a spec with a definition of done, a plan with milestones and acceptance criteria, a runbook and a running status log. Claude Code, Cursor, GitHub Copilot CLI and Gemini CLI now ship a plan mode, and Cursor says most of its own new features begin with the agent writing a plan.
Anthropic’s setup for long application builds opens with a planner agent that turns one to four sentences into a full product spec. The planner is told to stay at the level of product and high-level design, because errors in a detailed spec carry through to the rest of the build. The first version also had the building agent work in sprints. When a stronger model arrived, the author took the sprints out and kept the planner.
Human planning there follows the same pattern. On a recent Anthropic platform project, The Pragmatic Engineer reports, planning took the largest share of the effort.
What a bigger plan contains
A bigger plan covers a larger piece of work than a sprint, roughly the size we used to give a quarter. It states why we are doing this and which outcome counts, what has to stay true, where the interfaces are and how the result will be verified.
The last item carries as much weight as the rest. At Anthropic, by The Pragmatic Engineer’s account, verification takes longer than implementation, and in the Faros data review is where the time went. A plan with a definition of done and acceptance checks is what makes that checking manageable. A plan without them moves the queue from development to review.
A bigger plan also stays at the level of intent and design. The expensive hours of experienced people go into that document. The cheap hours of agents go into execution.
What stays from the agile years
The critics of plan-heavy development have a point, and several of them are the people who taught the industry to work iteratively. Kent Beck’s objection, relayed by Martin Fowler in January 2026, is that writing the whole specification first treats implementation as a phase in which nothing new gets learned. Birgitta Böckeler at Thoughtworks tried the specification tools and reported heavy workflows, tedious reviews of generated markdown and agents that don’t reliably follow the spec. François Zaninotto titled his critique “The Waterfall Strikes Back.”
They are right about frozen plans. Faster building does nothing about the oldest uncertainty in product work, which is that we don’t know what customers want until they have it in their hands. Cheap change is also an argument for committing late. So three habits from the agile years stay: testing the plan, delivering in small batches and keeping the plan open.
The plan gets tested before the organization commits to it. Winston Royce, whose 1970 paper contains the diagram we now call waterfall, called that approach “risky and invites failure” in the same paper and recommended doing it twice: a pilot version first, so that what reaches the customer is the second attempt. In 1970 that was unaffordable. A team with agents can build a throwaway version three times in a week and feed what it learned back into the plan. Bent Flyvbjerg, whose database of some 16,000 large projects shows 8.5 percent finishing on budget and on time, gives the same advice as “think slow, act fast”: do the trying-out during planning, where being wrong is cheap.
Delivery stays in small batches. DORA’s 2025 research lists working in small batches among the practices that amplify AI’s positive effect on product performance.
And the plan stays open to change. Bigger describes the scope a plan covers, and a plan of that scope can still change from week to week.
The fine print
METR’s numbers describe a 50 percent success rate on clean, well-specified software tasks, measured in human time. At 80 percent reliability the horizon for the same Opus model is a little over an hour, and METR calls its own measurements above 16 hours unreliable. Anthropic’s usage data from early 2026 put the median Claude Code turn at about 45 seconds. Long autonomous runs are still the exception. The Faros figures are vendor telemetry and show correlation. The lab examples come from the companies that sell the models, and at Anthropic engineers reportedly work without a token budget. My company is an Anthropic partner, so weigh my enthusiasm accordingly.
Agile taught us to plan the next step because the next step was what we could afford. The next mile has become affordable.
How much of your team’s planning still assumes that building is the slow part?
Tim Menzies, William Nichols, Forrest Shull, Lucas Layman, “Are Delayed Issues Harder to Resolve? Revisiting Cost-to-Fix of Defects throughout the Lifecycle,” 2016. arxiv.org/abs/1609.04886
I love talking to engineers. Not the business-media-trained kind who give you polished soundbites, but the ones who just tell you how it is. Recently, I got to sit down with senior engineers at OpenAI at the Pragmatic Summit in San Francisco. What they described wasn’t some futuristic vision. It’s how they work right now. And honestly, if someone had told me this six years ago, I would have thought they were making it up. Codex has gone from being a helpful coding tool to something the team literally treats as a colleague. Engineers assign it work in their task tracker, expect it to deliver results overnight, and start their mornings reviewing what it produced. More than a million developers now use Codex weekly, and usage has grown 5x since January. Inside OpenAI, nearly every engineer uses it and the teams merge 70% more pull requests each week compared to a year ago. But the numbers are just the surface. What’s really interesting is how it’s changing the day-to-day work. Let me break it down...
Workflows vs. Agents: The Distinction That Matters
I'll be honest: we're early. The agentic capabilities in Claude today are genuinely useful for my work, but they're not autonomous. Claude doesn't run my consulting business while I sleep. What it does is eliminate the grind. The repetitive tasks that used to eat two hours of every day and required zero expertise — just patience. That's what agents mean in practice right now.
One day, on-site, four to eight people from your company. We set up a fictional company, Novatech, to look like yours, with a working AI-native environment inside it. Your crew runs a real project in it. You get a one-page diagnosis of what has to change back at base.
By clicking “Accept All”, you agree to the storing of cookies on your device to enhance site navigation, analyze site usage and assist in our marketing efforts as outlined in our privacy policy.