Sequoia says agents handle 50% of work

The headline version of this story is a little off. Sequoia did not just announce that agents already do 50% of engineering work. What Sequoia actually has on the record is a prediction from GitHub CEO Thomas Dohmke, carried in a Sequoia podcast page, that AI agents could handle 50% of software engineering tasks by 2025 and 90% by 2028. Separately, Sequoia’s January 14, 2026 essay argued that coding agents have crossed a real capability threshold. Those are different claims — and that distinction matters. (sequoiacap.com) ### So what did Sequoia really say? On Sequoia’s site, the cleanest “50%” line comes from the page for a Training Data episode with Thomas Dohmke. The page says he expects AI agents to handle 50% of software engineering tasks by 2025 and 90% by 2028. That is not Sequoia presenting an audited market statistic. It is Sequoia amplifying a forecast from one of the most important people in AI coding tools. (sequoiacap([sequoiacap.com)e people treating it like news now? Because the surrounding environment changed fast. In Sequoia’s January 2026 essay “2026: This is AGI,” Pat Grady and Sonya Huang argue that long-horizon coding agents recently crossed a threshold where they can work autonomously for hours, fixing mistakes and deciding what to do next without constant human prompting. That makes an old forecast about “50% of tasks” feel(sequoiacap.com)t level. (sequoiacap.com) ### What does “50% of tasks” even mean? Basically, not “half of engineers are gone.” It means a growing share of the work inside software engineering can be delegated — writing code, debugging, testing, generating pull requests, and handling bounded tickets. Sequoia has been making this exact framing for a while. In its 2024 essay on AI for software engineering, it argued that the bottleneck is not just talent but time, and (sequoiacap.com)self. (sequoiacap.com) ### Why does coding go first? Because coding is unusually legible to machines. The task has clear artifacts — repos, tests, diffs, logs, pull requests. Success and failure are easier to score than in most knowledge work. Sequoia’s January essay explicitly says coding agents are the first concrete example of its functional definition of AGI, and OpenAI’s Codex team told Sequoia their agent can work independently in its own e(sequoiacap.com)le task descriptions. (sequoiacap.com) ### What about the Airtable and “30 Claude instances” line? I could not verify that specific claim from a primary source in the material available on the open web. There is plenty showing Airtable has gone hard on AI — including MCP support for Claude, ChatGPT, and Cursor, plus public discussion of an internal AI reorg — but not a solid primary-source confirmation here for “Howie Liu runs 30 Claude instances to ship code.” S(sequoiacap.com)view clip or transcript is produced. (support.airtable.com) ### What changes inside engineering teams? The job shifts upward a level. Less time on typing boilerplate. More time on task decomposition, tool access, evals, code review, reliability, and deciding which work should never be delegated. The catch is that agent output is only as good as the environment around it — tests, permissions, context, and rollback paths. Sequoia’s broader 2025 and 2026 writing keeps p(support.airtable.com)ne. (sequoiacap.com) ### Bottom line? The real story is not “Sequoia measured agents doing 50% of engineering today.” It’s that a once-aggressive forecast is colliding with much better coding agents, and the gap between prediction and practice is shrinking fast. (sequoiacap.com)