Every AI coding assistant of the last two years shipped the same feature: a plan mode. You describe what you want, the agent writes a plan, you approve it, and only then does it touch code. Claude Code has one. Codex grew annotation features around one. An entire category of startups was built on the idea that planning was about to become the most important part of programming with AI.
Last week a post called Plan mode is dead hit the Hacker News front page and sat there with 538 points and 469 comments. What makes it different from the usual hot take is the author. Ayman Nadeem did not watch plan mode from the outside. She built and launched an entire desktop coding app, Nuanced, with planning as its core feature. Then she watched it fail, and wrote the most precise postmortem of the planning-first workflow I have read.
Full disclosure: I have never used Nuanced and I am not connected to the author in any way. Everything about the product and the argument here comes from her essay, read alongside the Hacker News discussion. The opinion parts at the end are mine, and I run coding agents daily on my own projects, so I have a stake in this question.
The two jobs plan mode was doing
Nadeem's core insight is that plan modes were quietly doing two different jobs at once, and only one of them still exists.
Job one: specify instructions precisely enough for the agent to execute. In 2024 and early 2025, models needed this. You had to spell out the architecture, name the files, and describe the edge cases, or the agent would fill the gaps with guesses that hardened into code. A written plan was the contract.
Job two: help the human understand what they are building. The act of reviewing a plan forced you to think through decisions before they became five hundred lines of generated code across a dozen files.
Her claim, and the data point behind it is her own failed product, is that job one is rapidly becoming obsolete. Models got good at exploring a repository and making reasonable assumptions. Every decision the model can now reliably make by itself is one fewer decision that needs to be surfaced to you in a plan. Job two matters more than ever, but she argues plan modes are the wrong abstraction for it.
Four reasons the planning-first workflow failed
Nadeem names four reasons her own app did not deliver. Each one is worth sitting with, because they generalize far beyond one product.
She conflated planning with a plan. The thinking before the code is valuable. The long structured document that preserves the thinking turned out to be something nobody wanted. Early users had surprisingly little appetite for the spec her app generated. The process was the point. The artifact was not.
Models got really good. This is the competitive dynamic she names explicitly: model capability was competing with interface design. As models improved at understanding large codebases through context and memory, the need to instruct them explicitly shrank. A product whose value was upstream of the model found the model eating its value from below.
Nobody wants to read AI-generated text. This is the reason that stings. The specs Nuanced produced were long, accurate, and unreadable. Nadeem describes her own eyes glazing over at the pacing and over-structured nature of AI-generated writing. The team built a "Spec Tour" feature to walk users through the important parts instead of the whole document. Her verdict on that fix is brutal: if you need to generate a shorter summary to make the document usable, what was the point of the document?
They split a process that needs to be continuous. Nuanced's workflow ran in one direction: chat, disambiguate, generate spec, review spec, approve, implement, review code. Real thinking does not work that way. You understand part of the problem, try something, the result teaches you something, and you change your mind. Her app forced users to prematurely finish thinking before building could start, and once implementation began, going back felt like moving backward.
Existing plan modes were clunky before they were obsolete
Worth knowing where this frustration came from. Before building Nuanced, Nadeem rotated through the Claude Code CLI, then Conductor, then Codex. Her planning workflow inside those tools was cut and paste: shape a plan in chat, copy the good parts into a new message to revise them, repeat. The plan had no home. It lived as ephemeral text disappearing into backscroll as the conversation moved on.
That pain is real, and it is what the plan-mode-as-product thesis was responding to. She wanted plans to become a first-class primitive: a persistent document, grounded in the workflow, that carried decisions across the whole build. The diagnosis of the pain was right. The prescription, a dedicated planning interface, lost to the models and to plain chat.
What replaces plan, approve, execute
The old loop assumed the expensive mistake was writing the wrong code, so a human checkpoint before execution was worth its cost. Nadeem's argument is that agents changed the economics. They can act autonomously, test their own work, and revise their approach after inspecting the result. The loop that replaces the waterfall:
- Understand the task and the surrounding system.
- Act on a slice of it.
- Inspect the result, including running the tests.
- Clarify with the human only where the result surfaced a real decision.
- Adjust and act again.
There is still an enormous amount of planning happening inside that loop. What disappears is the document called "the plan" standing between thinking and doing as a mandatory gate. Nadeem's sharpest line is that her biggest mistake was turning the plan into an artifact instead of designing a process for improved human understanding.
She also concedes something plan mode defenders will enjoy: the simple chat interface, used as needed rather than by default, beat her purpose-built planning interface. The awkward part of every plan mode she tried was the meta-question it forced on you: does this task deserve planning? Deciding whether to switch modes was itself cognitive load, and remembering to flip the toggle was one more thing to get wrong.
The problem nobody has solved yet
Here is the part of the essay I think most readers will skim past, and it is the part that matters most for the next two years.
If plan mode is dead, what keeps you oriented? Nadeem's answer is honest: nobody knows yet. Understanding what an AI agent changed and why is still an open problem, and it gets dramatically harder as you scale from five parallel agents to hundreds. Keeping up cannot mean reading every conversation and interrogating each diff. Her proposal is that agents need to identify the fewest places where human attention has the greatest impact, and surface enough context to make that attention useful. That is a product-shaped hole in the market that nobody has filled.
This maps onto something I see running my own small fleet of agents: the review problem is no longer reading code, it is allocating attention. When one agent writes a 400-line diff, you read the diff. When six agents each write one, you need a system that tells you which of the six to actually look at. The tools are not there yet.
When I would still write a plan down
"Plan mode is dead" does not mean "never write anything down." Nadeem herself says the understanding job matters more than ever. So here is the checklist I actually use, offered as the save-worthy takeaway of this piece. Write a plan as a document when:
- The blast radius is production. Database migrations, auth changes, anything touching money. The cost of a wrong direction is still high enough that a checkpoint pays for itself.
- You cannot verify the result. If neither you nor the agent can test the outcome, the approve step is your only quality gate.
- The requirements are contested. Multiple stakeholders who disagree is a human problem. A document is how humans disagree productively.
- The work spans days, not minutes. Long-running work outlives chat context. Something durable has to carry the decisions across sessions.
For everything else, the loop wins: understand, act, inspect, clarify, adjust, act again.
The takeaway
Plan modes were scaffolding for a model capability gap, and the gap is closing. The builders who treated the plan as the product are watching that product evaporate. The lasting problem is the one Nadeem ends on: how humans maintain a coherent mental model of a system that machines are changing faster than humans can inspect. Solving that is worth more than any plan mode ever was.
I write about AI agents, developer tools, and backend engineering every week. Subscribe, it is free.
Have you stopped using plan mode, or do you still approve a plan before every agent run? Tell me what your loop looks like in the comments.
Top comments (0)