No category in AI has grown faster than "agents." Every product briefing now promises software that plans, acts and completes work on its own. The demos are genuinely impressive — and mostly irrelevant to how enterprises actually buy.
IT buyers do not purchase autonomy. They purchase reliability, auditability and a support contract. An agent that is right ninety percent of the time is not ninety percent useful; it is one hundred percent supervised. Until error rates fall into the range where human review becomes optional, most "agentic" deployments will remain pilots with generous marketing budgets.
That does not mean the category is empty. Coding assistants have already crossed the reliability threshold, which is precisely why they are the rare agent-adjacent product generating renewal revenue rather than press releases. The next winners will be the boring ones: agents with rollback, permissions, logs and a price tied to completed work rather than tokens burned.
The bubble will not burst so much as drain. When it does, the survivors will look less like magic and more like infrastructure.
Comments (0)
Log in to join the discussion
Log InNo comments yet