Delete your IDE
Chintan Turakhia is the Head of Engineering at Base and oversees AI initiatives within Coinbase. Over the past year, he’d been pushing agents deeper into every layer of product development. By early 2026, he felt something had shifted underneath them. Agents had become genuinely capable, and the workflows his team had built around the previous generation of tools were already holding them back. We spoke with Chintan about how he got there, and what he did next.
In January 2026, Chintan asked his entire engineering organisation to delete their IDEs and write zero lines of code. For two weeks, every engineer at Base would have to do their job without touching a code editor. The point was to figure out how work actually gets done when agents are doing most of it, not humans.
What makes this worth telling is everything that led to it, and the widening gap it reveals between companies that add AI to their existing workflows and companies that rebuild the workflows themselves.
At most workplaces, progress happens through asking questions. Unsure how something works? Slack a teammate. Notice a bug? Mention it in a channel. It’s the most natural way of working there is.
Asking costs nothing. Answering, on the other hand, is very expensive. It pulls someone out of whatever they were building, forces a context switch, and burns focus that’s hard to recover. Small teams can absorb this cost. At a company of thousands, with everyone taking the path of least resistance, it really starts to slow you down.
Standups, quarterly planning, quick syncs, status reports. These rituals exist to impose order on how people naturally communicate. They work, but they compress the time available for deep work. “I don’t think developers are slow,” Chintan says, “I think companies are slow.”
Chintan points to merged PRs per day per engineer as a useful proxy for the coordination headwind these rituals create. The number is almost always lower than anyone expects.
All the usual asking and answering and ritual-building creates a landscape where knowledge ends up scattered across people’s heads, Slack threads, meeting notes, and spreadsheets. Humans can navigate this mess. They can piece together context, read between the lines, and figure out what’s actually going on. Agents can’t, at least, not very well. They need a single, structured, accessible source of truth. The way most companies work today is a wall for agents.
Building the foundations
If agents need structured context to work autonomously, there has to be a single place where that context lives. That place is Linear.
Product requirements, designs, bug reports, shipping status, team structure. It all lives there. Chintan told his team from day one to treat Linear as the source of truth for everything, because agents depend on it. Linear has become the place where agents go to get their bearings before they start a task. “I’m not designing things for humans anymore,” he says. “I’m designing things for agents. And they need different things.”
With the rails in place, Base had to decide what rides on them. They started with Linear’s own agents, then built custom ones for work specific to them. The first is Claude Bot, which is a custom built agentic harness his team built that can be executed from Slack, Github, and Linear. It integrates directly into Linear through the Agent SDK, along with other tools.
Automations designed for the way teams naturally work
ClaudeBot and Linear both live inside Slack, which is where almost everything at Coinbase happens anyway. Chintan had no interest in changing that habit. He wanted to turn those conversations into actionable work, and stop the stream of questions from pulling people out of deep focus.
When someone flags a bug in a channel, an automated workflow picks up the message and creates an issue in Linear. The issue gets automatically labelled, sized, and assigned. Made possible because Linear has all the context around their team structure, plans, and related work. No one has to read the thread, interpret what was said, or manually create an issue. The conversation just becomes work.
Because the new issue is a bug, it gets automatically assigned to Coinbase’s agent, Claude Bot. With access to their codebase, company knowledge, and all the context in Linear, Claude Bot works on a fix, drafts a PR, and pushes it back to the team in Slack for review.
The other side of it is the questions. Someone asks “what’s the status of this feature?” or “when is GA?" and instead of pulling a teammate away from whatever they’re building, the question is picked up directly by an agent. Linear handles anything related to project status, timelines, and shipping. Claude Bot handles the code specific stuff, like debugging a service or figuring out why a particular endpoint is failing.
Engineers stay in flow. Work gets created from conversations that would have otherwise been forgotten by the next morning. And the rituals that existed to manage all of this coordination (the standups, the syncs, the status reports) start to feel like they’re solving a problem that isn’t really there anymore.
Getting people to adopt new tools is usually a project in itself. Training programs, rollout plans, and the dreaded change management. None of that was necessary here, because engineers didn’t need to learn a new behaviour. You ask a question in Slack, just like before. The only difference is who answers. And because this happens in public channels, every time someone sees an agent respond with something useful, it reinforces the habit for everyone else watching.
Making it stick
Some engineers were already pushing the boundaries of what agents could do. Others were sceptical, or just hadn’t felt the urgency yet. Chintan’s objective was to push everyone up that spectrum, so he introduced something called “Speedruns”.
Almost weekly, one or two power users demo a new workflow or capability to the entire organisation. Showing, rather than telling. Afterwards, everyone on the call did the workflow, live. It’s informal and slightly chaotic, but it sets the cultural direction.
The first one happened in early 2025. Chintan ran it from his phone in the back of an Uber after landing in New York, walking about 120 engineers through new workflows around agents writing tests and drafting PRs. They put up 80 pull requests in a single sitting and crashed GitHub. By mid-year, CEO Brian Armstrong was attending. In one session, over 500 PRs went up in fifteen minutes. They’ve taken GitHub down four or five times now (which Chintan admitted with a hint of pride).
Beyond the speedruns, there are Slack channels dedicated to sharing AI wins and automations, and an Applied AI Award that Brian presents at quarterly town halls. All of it is designed to make new ways of working visible across the company, so that breakthroughs discovered by one team don’t stay siloed there.
What continuous development at Coinbase looks like
Once the culture took hold, new patterns started to emerge. Engineers began waking up, reviewing the pull requests agents had generated overnight, then spinning off ten to fifteen new agents to work throughout the day. They’d go heads down on the complicated stuff themselves. At the end of the day, they’d review again, launch another batch to run overnight, and wake up to a fresh set of results.
Development at Base is becoming continuous, not because people are working around the clock but because agents are. Chintan tracks autonomous operation time as a key metric, how many minutes an agent can run without a human stepping in. That number keeps climbing.
When Chintan talks to his peers at other large companies, he sees a very different pace. Most haven’t felt the urgency yet. They’re still evaluating which coding agent to adopt. Which is a good first step, but from what Chintan has learned, adding AI to existing workflows will only ever get you incremental gains. The step change comes from rebuilding the workflows themselves.
The pause in January was designed to do just that. Weeks later, those workflows looked different again. They’ll probably look different again in six months. At Coinbase, every conversation now starts with the same question. “Why can’t an agent do this, and how do we make it possible?”.