Do we still need an IDE: how I split work with an AI agent and what I keep for myself

JetBrains put out three texts in one week: an essay on why an IDE is needed in the age of agents, the release of Air Governance for managing agents and spend, and figures on which work developers are willing to hand to AI. GitLab proposed three tasks for an agent, Simon Willison shared an obituary for programming as puzzle-solving, and Latent Space described mods for Claude Code. For a business owner who orders development, there's one question: what in the contractor's work does the model do, and what does he answer for himself.

What came out

JetBrains on October 20 published the essay “Why We Still Need an IDE.” The author compares agentic development to flying on instruments: the flight computer flies the plane, and the pilot sees the whole picture and is responsible for the flight. In the same way, an engineer is responsible for what goes out in a release, even if an agent wrote most of the code. The author himself built a browser extension in an unfamiliar stack: the agent did the mechanical part, and to understand and fix it he went into the IDE.

JetBrains on October 22 announced general availability of Air Governance: one console where a company sets the allowed agents, models and providers, puts spending limits on people and teams, and configures MCP servers at the company level. Their reason is their own: JetBrains' AI spend grew about tenfold in the first half of 2026, and the first consolidated report across vendors took four days to put together by hand.

JetBrains on October 19 showed how its surveys have moved: willingness to hand application logic to AI grew from 18% in April to June 2025 to 50% a year later, tests from 50% to 75%, refactoring from 38% to 60%. 43% are willing to hand over code review, against 33% a year ago. 54% are not willing to hand over debugging, against 76% a year ago.

GitLab on October 19 proposed three tasks for an agent on the Duo platform: a first pass of review on a merge request, analysis of a failed pipeline with a suggested fix, and answering questions about the project in chat.

Simon Willison on October 19 picked out Jeremy Scheff's piece “My obituary for programming” about what's going away: the work was paid solving of logic puzzles, and the trickier the puzzle, the better it paid. Willison adds that anyone who cares about building working systems is glad to have the models.

Latent Space on October 21 described Claude Mods, plugins in TypeScript or JavaScript that change the behavior and interface of Claude Code. An example from the article: the built-in You Should Know mod noticed that the file the agent was editing, and about 45 other files, held uncommitted work, and suggested saving it first.

Example: parts calculation for a manufacturer

Take a manufacturer where the makeup of a product is described in Excel specifications, and the parts and materials for an order are calculated by hand. What's expensive in that task is the rules, not the code: how parts are grouped by the processing they need, what to do when an item isn't in the catalog, when two orders can go to a subcontractor together. The technical director knows these rules, and they have to be drawn out of him and agreed on. The agent will write the function, but it won't go and talk to the technical director.

AI speeds up the part where it's already clear what to do. The part where you have to decide what to do stays with a person, and that's what the client pays for.

How it works for me

I started programming in 2017, when my companies needed systems that didn't exist off the shelf, long before generative tools. Today AI has sped up my development several times over: routine, tests, debugging. Architecture and the fundamental decisions I make myself, and I check what the tools produce.

The manufacturing company shows it best. From February 2024 the technical director and I worked on the specification, active development started around April, I handed over a working program in July or August and refined it through the end of the year. The program calculates one order in 65 seconds on average instead of 4 to 5 hours. Today an agent would remove part of the barrier of an unfamiliar stack, as it did for the JetBrains author, but it wouldn't cancel the months of agreeing on rules with the technical director.

At the PDR company in the US, where I've been the technical partner since 2021, we build the React Native super app together with the company's developers, and the owner now does development with us himself. With the JetBrains numbers I mostly agree: tests and refactoring I hand over gladly, debugging where I can re-check it, and review of changes I keep for myself.

What I would do

Ask your contractor what the model does in his work and how he checks it. A normal answer is specific: tests, drafts, searching through other people's code, with architecture and release staying with a person.

Agree on who owns the result and where it lives. With me, the code, the database, the server and the access credentials belong to the client, and that doesn't depend on which tool helped write the code.

If you have your own team, look at where the money on agents and models goes. JetBrains' four days for a manual report show that without shared accounting, AI costs grow unnoticed.

Leave review to a person, even if an agent does the first pass, as at GitLab. Hand over debugging where there's a test that will show things got better. Skip the argument about whether an IDE is needed: what you need is a place where the whole change is visible before it reaches the server.

Related reading

On rules, statuses and permissions that both people and AI work by: the article Business logic: rules, statuses and permissions. Services: AI in a process where the result can be verified and the business as one system. Projects: Technical partner to a PDR company in the US since 2021 and Order calculation in 65 seconds instead of 4 to 5 hours. Nearby are my reviews on what I let an AI agent do on a production server and on sandboxes for agents and releasing kononov.dev, and earlier I wrote about where quickly prepared work gets stuck after AI. If you're choosing a contractor or building AI into your team's work, tell me where you are now. Discuss your task.