Exclusive: Cognition bets on ‘Devin Bots’ coding agents in challenge to Anthropic’s Duet and OpenAI’s Dots

Exclusive: Cognition bets on ‘Devin Bots’ coding agents in challenge to Anthropic’s Duet and OpenAI’s Dots

Cognition’s AI coder just got promoted. The startup, currently valued at $48 billion, helped kick off the race to build coding agents with its viral product Devin. Now it is trying to get back to the front of the pack by debuting Devin Bots, a persistent coding tool that acts as a manager for teams of Devin agents that can manage a complex project over days or even weeks. In the past few months, several AI companies have rolled out always-on agents that are designed to be general-purpose, but can be used for coding tasks. OpenAI has unveiled Dots, an always-on proactive agent that integrates with its Codex product to run teams of agents handling software development tasks. Anthropic has unveiled a similar product called Duet that can integrate with its Claude Code product.But Cognition’s Devin Bots are some of the first to be designed specifically for software development. Tell a Bot to build an Android version of your iPhone app, and it will divide up the work, assign it to Devin agents and keep the Android app updated as the iPhone version changes. “Self-driving software is, I think, the direction that it’s going,” CEO Scott Wu told Fortune at the company’s office in San Francisco’s SoMa neighborhood. “But of course, it’s always up to the human where we’re driving to.” Devin Bots is Wu’s bet that the smarter AI models get, the more the product built around them matters. His argument addresses a broader Silicon Valley debate over whether advances in AI models will squeeze out the companies building products on top of them. Wu disagrees with the idea that AI will eventually become so capable that, as he put it, “nothing else will matter at all.” When Cognition unveiled Devin in March 2024, calling it the first AI software engineer, it was among the early startups trying to get AI to write and ship code on its own. The launch video went viral. Two and a half years later, coding is one of AI’s biggest businesses. OpenAI and Anthropic have both put heavy resources behind their coding agents. In August, SpaceX acquired Anysphere, the parent company of AI coding tool Cursor, in a transaction that valued the company at $60 billion.For Cognition, the model companies are both suppliers and competitors. Their advances can make Devin more capable, while improvements to Codex and Claude Code could give customers less reason to pay for Devin as well. Wu argues that as models become more capable, engineers need more ways to direct them and inspect the results. GitHub Copilot began by suggesting code that programmers could accept or ignore. An engineer working with a Devin Bot, he said, might tell it what to optimize, how to use the team’s project board and when to send updates, or ask for a video of it testing its work. He compares this to how colleagues work together on a team. “Even here in this office, as you can see, we have whiteboards, we have scratch paper,” he said. Engineers discuss requirements, look at diagrams and try what has been built. In his view, the more AI can do, the more ways people need to communicate with it. A Devin Bot joins a team’s Slack as its own member, remembers context across weeks and contacts people when it is blocked or needs a decision, according to the company. Several colleagues can work with the same Bot, including over iMessage. The product incorporates technology from the Interaction Company, the startup behind AI assistant Poke, which Cognition acquired in July. OpenAI and Anthropic are also building agents that work across connected apps and act proactively. Wu said Cognition’s one advantage is that it isn’t tied to a single lab’s models. A Bot selects suitable models for planning, design, computer use and other parts of a project, using faster, cheaper ones for routine code. “Developers will especially want to optimize cost as they use products like Devin Bots that do way more work than before,” he said. Each Bot can be given its own budget, measured against the results it delivers. Cognition also trains its own models, which Wu said are designed to perform software engineering and code review at lower cost than frontier models. Devin combines them with models from outside labs. Cognition said usage of Bots across dozens of early-access customers, including Decagon, Fireworks AI and Reducto, grew tenfold in two weeks. The company announced its annualized revenue run rate passed $1 billion in September. Its existing Devin product is used by 14,000 engineers at Citigroup, CEO Jane Fraser said on the bank’s Oct. 13 earnings call. She said Devin made in-house development more efficient, adding, “We’ve been in-sourcing work and have reduced our tech vendor spend by 30% year-over-year as a consequence.” Handing a Bot a project also means deciding which systems it can access and which changes it can make without approval. Cognition runs Bots on separate cloud computers, keeping them away from passwords and keys stored on engineers’ laptops, Wu said. Administrators set which code repositories and tools each Bot can access. Network controls are intended to prevent unauthorized data transfers, including if a model follows malicious instructions hidden in content it reads. Wu said Bots run their code through the same tests used for human-written code. Teams decide which changes require human review and can configure low-risk edits, such as changing a button’s color, for automatic approval and merging. The safeguards, he said, rely on rules companies set and monitor, “rather than relying on models themselves being smart enough to recognize a risky or malicious prompt.” Wu said his aim is to keep making Devin more autonomous as models improve. He wants Cognition positioned so that “we are happy when the frontier capabilities continue to improve, not sad.”

Original Source

Read the full article at Fortune →

KhanList aggregates and links to publicly available news content. We do not host full articles from third-party sources. Always verify important information with original sources.