When coding agents can write an implementation in minutes, the engineer’s job shifts from typing code to deciding what should exist, what “acceptable” means, and whether users ended up better off. That is the argument in Kent C. Dodds’ essay of September 24, 2026, which condenses into one line: “Agents produce output. You own the outcome.”
This article walks through that argument, the working method he describes, and the one concrete example he uses. It treats his claims as one practitioner’s argument and experience, not as measured evidence about the labor market or about productivity.
As an Amazon Associate I earn from qualifying purchases.
Output versus outcome
The essay turns on one distinction. Code, pull requests and merged changes are output. A useful effect for real users is the outcome. An agent can generate the first at high speed. Someone still has to judge whether that output was worth producing and whether it helped anyone.
Free tools Windows power users keep installed
One-click scans. No signup required.
Dodds frames the career worry as “what’s left for us?” and the practical version as “what should you be spending your time on?” His answer is judgment and accountability. Treat it as his thesis. The essay offers no study or statistic showing how this plays out across the industry, and this article adds none.
#1 Best Overall
The working method
1. Reach shared understanding before the agent writes anything
Agree first on the problem, who experiences it, what they are trying to do, and what “done” looks like. Without this, an agent can build something polished that solves the wrong problem.
2. Ask for options, not just an implementation
Dodds asks the agent for candidate approaches, each with rough effort, trade-offs, reversibility and the agent’s own recommendation. Then he makes the call. The agent supplies breadth, and the human keeps the decision.
3. Scale scrutiny to reversibility and risk
Not every decision deserves the same care. The essay treats API names, data shapes and pricing as consequential, effectively one-way choices, because users and integrations come to depend on them. Choices that are easy to undo can move faster. Put your review time on the first group.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minute4. Build the environment that makes delegation safe
Dodds says he did not read the diff line by line for the feature he describes. He relied on setup done beforehand. The essay names these parts of that environment:
Rank #3
- trusted automated gates that catch regressions;
- isolation of secrets;
- self-healing automations;
- metrics that show what is happening in production;
- an understanding of unit economics, meaning what each feature costs to run.
The point is that skipping line-by-line review is only defensible when these controls exist. They are what make trust reasonable.
The example: testing webhooks in Kody
Kody is Dodds’ SaaS product for storing durable software that agents can share and reuse. He says it is meant to work alongside agent platforms such as Claude, Cursor, Devin and OpenAI. He wanted users to be able to test a webhook handler before real traffic arrives.
What he shipped
A synthetic dispatch. It invokes the webhook handler with a supplied fixture, through the product’s MCP and API surface.
What he deferred
He postponed a UI button and a full end-to-end ingress dry run. The stated reason is that the smaller slice answers the immediate question, which is whether the handler works, without the cost of building everything around it.
Best Value
Why he avoided the name “dry run”
A dry run suggests nothing real happens. Here the handler can still call real services and cause side effects. Calling it a dry run would have misled users, so he chose a more honest name. He also notes that synthetic runs consume compute and count toward usage, so cost is something to state, not hide.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How the decision compares across options
| Option | Answers the user’s question? | Effort | Side-effect risk | Reversibility |
|---|---|---|---|---|
| Synthetic dispatch with a fixture (shipped) | Yes: does the handler work? | Smaller slice | Real; the handler can still call real services | Not stated in the essay |
| UI button (deferred) | Same question, easier access | More work | Not stated | Not stated |
| Full end-to-end ingress dry run (deferred) | Broader: tests the whole path | Largest | Not stated | Not stated |
The essay gives qualitative reasoning, not scored numbers, so the table records only what it supports. The pattern still carries over: weigh effort, the user need answered, trade-offs, reversibility, side-effect risk and compute cost, then pick the smallest slice that answers the question.
What this asks of you
- Name things honestly. Labels are product decisions. A misleading one becomes the user’s mistake.
- State costs and risks in the product. If a feature uses compute or touches live services, say so.
- Ship the smallest useful slice. Agents make large builds cheap, which makes it tempting to build the whole thing. The question to ask is what the user needs answered now.
- Invest in gates before you delegate. Reduced review is a result of that investment, not a shortcut around it.
Try it on one feature this week
The essay closes by asking readers to pick a single feature and apply the process. A practical version:
- Write down the problem, who has it, what they are trying to do, and what “done” means.
- Ask your agent for two or three approaches with effort, trade-offs, reversibility and a recommendation.
- Mark which choices are hard to reverse (names, data shapes, pricing) and review those closely.
- Choose the smallest slice that answers the user’s question, and list what you are deferring.
- Check that your gates and secret isolation would catch a bad change before you let the agent run.
- After release, look at whether users benefited, not only whether it merged.
Limits of the argument
The essay comes from one builder working on his own product. It relates the ideas to a talk at a Distillery Tech Night with the React Buenos Aires community. It does not test the claim that judgment stays scarce, and it does not compare agent tools. It is a useful model of how to work, not proof of where the profession is heading.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
