Analysis
OpenAI continues to expand its agent lineup across browsing, coding, research and workflow automation, and the question TechCrunch raises is whether ordinary users will adopt them at anything like the rate they adopted chat. Pulse has previously covered OpenAI's product expansion from a single chat interface into a portfolio of agent products.
OpenAI, founded in 2015 by Sam Altman, Greg Brockman, Ilya Sutskever and others, has shipped agent products under names including Operator for browser tasks, Deep Research for multi-step web investigation, and Codex for autonomous coding -- each aimed at a category with its own well-funded startups. Operator competes with Adept's browser-automation approach; Deep Research overlaps with what Perplexity and a wave of research-agent startups have built; Codex competes directly with Cognition's Devin and GitHub Copilot's agentic modes. Sierra, founded by former Salesforce co-CEO Bret Taylor, has staked its entire business on customer-service agents being a defensible category rather than a feature OpenAI eventually ships natively.
“Chat succeeded because it demanded nothing new from the user -- type a question, read an answer, discard it if wrong.”
The distinction matters. Chat succeeded because it demanded nothing new from the user -- type a question, read an answer, discard it if wrong. An agent asks for something harder: permission to act, and trust that it will not book the wrong flight or email the wrong client. The failure mode of a bad chat response is a wasted minute. The failure mode of a bad agent action is a mess someone has to clean up.
Enterprise evidence suggests the constraint is not model capability. VentureBeat has reported that agent reliability tracks the quality of the documents and systems behind them, and that the deployments working best are the ones limiting how much agents do without a human check. Both findings point the same direction: agents work where the surrounding process is clean, which describes a minority of real companies.
For founders, OpenAI shipping agents horizontally across categories is the platform-risk conversation from every prior era, with a shorter clock. The defensible positions are the ones OpenAI structurally cannot occupy -- proprietary data, regulated workflows, systems of record, liability the customer needs someone to carry. General-purpose competence is not a moat when the model vendor also ships the application.
