Breaking News: IBM's AI Coding Tool Bob Goes Live with 80,000 Internal Users
IBM has announced the rollout of its AI-powered coding assistant, Bob, now used internally by over 80,000 developers. The tool, unveiled this week, represents the culmination of more than two decades of research by Neel Sundaresan, the company's General Manager of Automation and AI for IBM Software.
Sundaresan, who also served as a founding engineer of Microsoft's GitHub Copilot, emphasizes that Bob is built on a philosophy radically different from many current AI coding products. “Most AI coding today is like taking your Ferrari to buy milk,” he told The New Stack. “It's overkill for simple tasks and misses the real friction points.”
The Philosophy Behind Bob
Sundaresan argues that the user experience of a coding tool matters more than the sophistication of the underlying model. “Coding is an analytical task, different from shopping online,” he explains. “If the system makes a wrong recommendation or interferes with my thought process, that matters—a lot. You can have a better model and a worse product if you get the surface wrong.”
Bob is designed to minimize friction at precise moments in a developer's workflow. Rather than generating large blocks of code, it focuses on speeding up specific, repetitive tasks—an approach rooted in Sundaresan's earliest work.
Background: Two Decades of Developer Productivity Research
Sundaresan’s journey began in 2000, long before transformers or large language models existed. He built a recommender system for API calls, noticing that “30% of developer code is API calls.” The goal was not code generation but surfacing the right function at the right moment—a search-ranking problem applied to autocomplete.
Developers loved that early system. That signal—that reducing friction at a small, specific moment produced outsized satisfaction—shaped Sundaresan’s thinking for the next 24 years. He watched model architectures evolve from Long Short-Term Memory to transformers to GPT, but his team had already identified the core problem. “If you go back and look at our publications, we have publications in all of this,” he says. “Every paper would say, here’s the model… but the models just weren’t powerful enough yet.”
What This Means for the AI Coding Landscape
Bob’s internal rollout challenges the prevailing narrative that bigger models and broader code generation are the only path forward. Sundaresan insists that the user experience is orthogonal to AI capabilities. A tool can have a world-class model but fail if it disrupts the developer’s natural flow.
For enterprises evaluating AI coding assistants, Bob signals a shift toward targeted, low-friction integrations. “It’s not about generating entire functions,” Sundaresan notes. “It’s about knowing when to suggest, when to stay silent, and always respecting the developer’s thought process.” IBM plans to release Bob more broadly later this year, potentially reshaping expectations for what an AI coding tool should actually do.