Linear CEO: AI coding agents just add bandwidth, not replace engineers

Linear CEO Karri Saarinen just published a lengthy take on the real state of AI coding tools, and his verdict is refreshingly grounded. Agentic coding — where an AI agent plans, edits, debugs, and commits code — is already widespread. Linear's own data backs that up: most paid workspaces have installed a code agent, usage has grown more than 5x in a few months, and Linear's own cloud agent now fixes over 1,000 issues per month (and climbing fast).

This shift mirrors OpenAI's recent decision to merge its standalone Codex model into GPT-5.4. But Saarinen is measured about what it all means. Privately, he says, almost nobody claims agents write 100% of code, and you rarely hear about real companies running fully independent agent clusters at scale. Engineers still have to provide direction, constraints, and judgment — they typically manage a few local agents simultaneously and maybe let a cloud agent handle grunt work in the background.

He sums up AI's main impact on programming as "increasing bandwidth." Problems that used to be too small, too annoying, or too time-consuming are now easier to knock out. But truly hard problems haven't been accelerated by the same magnitude — they still require understanding systems, making tradeoffs, and deciding what should exist in the first place. DeepSeek's recent API cache price cut makes that bandwidth even cheaper, but Saarinen argues AI won't make planning obsolete. When building costs drop, it gets easier to build the wrong thing, too. Teams still need to decide what matters most — just with shorter planning cycles and more room for experimentation.