Anthropic Engineer: 99% of People Treat the Chat Window as an Agent

A senior engineer from Anthropic (the company behind Claude) said in a talk:

“Prompting is no longer the job. 99% of people still open one chat and call it agents.”

Writing prompts is no longer the main job. 99% of people still just open one chat window and call it an agent.

The remark was posted on X along with a video:

The comments yielded a lot of substance.

One commenter said prompt engineering is turning into “context engineering”: from figuring out how to ask the model to designing the entire context the model can see.

Another said that going from a single chat window to an agent comes down to two things: each agent handles only one clearly defined task, and there is a checkpoint it must pass before handoff. They then asked: what task would you hand off to a second agent?

Someone joked: my agents are three chat tabs plus a lot of hope.

Another commenter added that for multiple agents to work, each needs a narrow task, its own tools, and a checkpoint to pass before handoff. Otherwise, you just get five chat windows arguing with each other faster.

Someone summed up the current work as: write specs tightly enough that the agent has no room for creativity.

Someone teased Anthropic employees: walk in and have Claude improve itself, all while making over $300k a year.

These replies all explain the same thing: the key to agents is task decomposition—breaking one big thing into a process with boundaries, checkpoints, and handoffs. Writing prompts is just one part of it, and not even the hardest part.

The trouble is, it’s easy to fool yourself. Opening a new tab and calling it an agent costs nothing. But that’s just giving a chat window a new name.