The next developer workflow needs more than an agent, it needs you.

Kat Hughes — Fri, 30 Oct 2026 20:45:44 +0000

It's awesome that agents can crank out code faster than ever, but who actually understands the change? Who reviews it? And how are you supposed to keep track when you've got five agents running at once?

That's what we dove into at our latest Community Office Hours with Jason Torres, a developer advocate for WebStorm at JetBrains.
Joining alongside me was Ben Potter, Chris Boone from Coder.

One of the first things Jason asked was whether he was allowed to swear. We're profanity friendly, so fair warning before you watch without your headphones!

Agents get you 80% of the way, but the last 20% is on you.

In his Commit Your Code 26 talk, Jason made the case that agents get you roughly 80% of the way, so we asked him what he thought the other 20% was. He said it depends on the project, and a lot of it comes back to the prompt. Sometimes it's honestly easier to write the code yourself than to explain it to an agent.

And agents only know what you tell them:

"They don't make assumptions. They don't infer things. And if they do, nine times out of 10, it's wrong."

If you don't understand what you're building, you won't know what to fix. Jason compared it to a plumber who actually knows the tool and isn't "just spinning a wrench and praying for the best."

Faster code just moves the bottleneck

At Coder, AI-assisted contributions completely changed code review. Our PR queue used to hold around 10 pull requests at a time being worked on by 50 engineers. Now about 100 people across the company submit PRs with a decent amount of them non-engineers and review is the new bottleneck.

Jason's take was basically accountability:

"You're still on the hook for that work."

An agent can't get fired. If you open the PR, you need to understand it and be able to fix it. He even floated PR validation as a future job title, then admitted it sounds like a terrible job.

Over at JetBrains, he said AI hasn't caused many bottlenecks like this yet. They use their own tools, and with tests and diffs right there in the IDE, it's easy to double-check before opening a PR. Their headache is figuring out how to build and message the product itself.

Not everyone bought it, though. A viewer in our Discord pushed back with a sentiment of nobody double-checks every prompt in real life, and if you have to babysit the agent, what's the point? Honestly, fair. Jason agreed it isn't sustainable right now. "It's meant to be iterated on," he said. Agents will need less review as they get better, and we'll get better at prompting them. But some due diligence still stays with the human who's getting paid to do the work.

Managing Multiple Agents & Orchestration

We asked what workflow he uses to get the best output out of an agent. Jason said he hasn't found a perfect workflow yet, but skills have worked best for him so far. His animated video clips kept drifting in style from one to the next, so he wrote a design Markdown file and a skill file that the agent follows together. He also added steps so teammates can reproduce the same results.

Some engineers at Coder are juggling five to ten agents at once. Jason said the disconnect happens when that work is scattered across tools with one task in Claude desktop, another in a CLI, and then you totally forget the second desktop task while you're off in the terminal.

His own setup keeps the agent, terminal, test runs, and diffs side by side in the IDE. JetBrains is bringing multi-agent sessions into Air so they live in one window, and it's working on letting you switch providers mid-conversation (say, from a JetBrains AI key to an OpenAI key) without losing context.

This feels like a natural place for Coder and JetBrains to team up. Ben mentioned that our DevRel team could build a solid JetBrains Air integration and some agentic workflow examples across both ecosystems. To be clear, that was an idea we tossed around on the call, not an announcement. We like to do a lot of ideation on the Dev Rel team.

Will all these new AI innovations pull people out of the IDE?

With so many new AI innovations launching lately, specifically with Dots / Muse / Duet, we asked if developers are going to stop opening their IDEs.

Jason doesn't believe so. In his view, those tools aren't really aimed at "the bubble we live in." They're helpful for people who don't even know what ChatGPT is yet. He told a story about meeting a union electrician with 20 years in the trade who had no idea what Jason did for a living. The tools that help them day to day could be a huge deal, but for now, he doesn't think that would replace the IDE.

Software Factories

Near the end, we asked the chat about software factories, where agents write, review, and QA the code and humans only show up at the start and end.

Jason brought up the It's Always Sunny in Philadelphia episode where the gang visits a factory and finds one guy in a lab coat watching the machines. He thinks we could eventually land somewhere near that, just not that extreme. For now, people can still spot AI-built work fairly easily, right down to the Tailwind blue-and-purple cards that, as he put it, "all the agents think we want to see." Until that gap closes, be it 5 years or 1 year with how fast AI innovation is going, humans will still thankfully need to provide input to differeniate themselves while using AI.

He also brought up cost with remembering a big company slashing employees' token budgets after spending got out of hand. An agent running for an hour isn't automatically cheaper than a person. Again, AI isn't meant to replace folks, but so more managing their AI agents.

Keep the conversation going

Before we ended, Jason's left with some parting advice for the community with

"Keep reviewing your code, keep building."

We are super thankful Jason was able to join us for our office hours.

Our Office Hours happens every two weeks, we'll advertise if we're going to be bringing on a guest, but other than that, it's mainly the Dev Rel team vibing with the community.
How are y'all handling review and context switching with agents? Drop it in the comments or come hang out in our Discord.

The next Office Hours is Friday, November 6, at 10 AM CST (16:00 UTC). Set a reminder on YouTube. See you there!