Cognition's Jeff Wang on demoting the IDE, with Richard Liu from Anthropic

How AI coding agents are changing software development Zero-Shot Learning is a podcast about how AI gets built, secured, and deployed. Hosted by Nancy Wang, 1Password CTO, and Dev Tagare, Senior Director of Engineering at Google, it’s a builder’s view of the architecture and the complex decisions it takes to ship with AI. This episode features guest co-host Richard Liu, Head of API Products at Anthropic, who sits in for Dev. The IDE has been the center of software development for decades, but Jeff Wang thinks its time in the spotlight is ending. On Zero-Shot Learning, the President of New Enterprise at Cognition described how his team’s workflows are shifting from manually writing every change to delegating work and verifying the results of AI coding agents. After leading Windsurf and now working with Devin at Cognition, Jeff has seen developers across industries explore how to implement AI. From interactively collaborating with coding agents to deploying long-running agents in the cloud, he has seen what agents change inside the editor and what they require outside of it. If the question is what’s the most effective way to ship, Jeff isn’t betting on the IDE. Offloading undesirable work to AI coding agents For Jeff’s team, AI coding agents prove useful for the work no one volunteers for. Agents respond to event-triggered tasks like reproducing bugs, remediating vulnerabilities, and repairing CI failures, where they effectively reduce noise that distracts developers from planned work. “If you go into any engineering organization, you don’t want to take away the things they want to do,” Jeff says. “You want to take away the things people don’t want to do. You ask everybody in the development team, ‘Hey, who wants to replicate this bug?’ Nobody is going to raise their hand.” He says these agents now account for roughly 40% of the workload at Cognition. Cognition also works with enterprise customers using Devin. At some large banks, Jeff says, Devin automatically fixes 70% of their vulnerabilities. By agentifying these workflows, Jeff says Cognition merged roughly 700% more pull requests over six months while increasing headcount by only 10%. With agents doing the grunt work, developers don’t need to be the sole operator of each task, fundamentally transforming how each team member spends their day from coding to orchestrating operators. “You might queue up a bunch of agents in the morning, go get lunch, and come back and unblock the ones that are stuck,” he said. In the interview, Richard Liu, Head of API Products at Anthropic, recounts that across industries, the average developers typically spend eight to 15% of their day doing hands-on coding. The rest goes to coordination, research, and meetings. With such powerful agentic systems at play, the IDE no longer defines the software development workflow. Instead, engineers decide which work an agent can take on, where it is blocked, and when the result is ready to verify.​ What does an agent-centered development environment require? An editor gives developers a single console to work from, but agentic systems work across systems and environments. While one agent might investigate a bug while another runs tests and a third waits for access or context, teams need a way to see each agent in action. OpenAI calls this supporting system a harness: the tools, application context, and feedback loops that help agents work. Cursor’s 2.0 release takes a similar approach from the product side, putting multiple agents at the center of the interface and recognizing that review and testing become harder when agents work in parallel. Still, these are only the start of a larger workflow change. Jeff describes a highly accessible version of that environment in Windsurf 2.0, where developers oversee a Kanban board showing what agents are doing and where they are blocked. A team member can start several tasks, return later, and support the agents that need help. For larger infrastructure work, agents can also move between local and remote environments. A developer might explore an idea locally, make a plan, and then hand the defined task to Devin in the cloud. Jeff says those remote agents run with the dependencies and data they need, allowing work to continue without keeping a developer at the keyboard. Once work is moving across agents and environments, the issue is to determine what is required for the agent to prove the work is done. Agents can differ both in where they run and in the authority they receive. 1Password’s agent architecture framework explores the distinction between local and remote deployment, as well as delegated, bounded, and autonomous authority. How do developers verify work completed by AI coding agents? When Nancy asked what the new standard primitive would be, Jeff answered, “Probably the ability to prove something is done.” With AI coding agents, frameworks to validate completed work must be designed before the work begins. Jeff points to the bene

Excerpt shown under the publisher's RSS terms · 99 words · full text and figures are on the publisher's site