Anthropic Engineer Builds an Agent in 10 Minutes; the Comments Section Is Doing Different Math

A 10-minute terminal screen recording. When the screen first opens, it is a blank terminal; by the end, an AI Agent system is already running: a single Agent handles execution, automatically checks after running, makes changes after checking, and finally several Agents work together.

The original post is from X user Andy L and is said to be a live demo by an Anthropic engineer. The post mentions a claim: 95% of engineers inside Anthropic use a similar workflow. This number cannot be independently verified, so keep it in mind for now.

The demo itself is fast. From a single Agent to multiple Agents, the closed loop of execution, checking, and improvement in between really does look smooth.

But after scrolling through all 15 replies, you realize the focus has basically gone off track. Few people ask how it was built in those 10 minutes; most are talking about what happens after it is built.

The checking step is the easiest to skip

The most frequent word in the replies is "checking."

User king says that in this chain, the checking step is the easiest to skip. User ReignSwords puts it more directly: building the system takes 10 minutes, writing the checking criteria takes 1 hour. He also mentions a detail: you cannot let the model doing the work audit itself in the checking step; you need a different perspective, for example another model plus clear acceptance criteria, otherwise it is just praising itself in different words.

Where User A-Suan gets stuck is not in building either, but in how to write the judgment criteria for the checking step.

The consensus among several users is roughly: when the same model self-checks, most of the time it is just rewriting the wording. Hand the checking to another Agent and give it tests or specs to compare against, and the result is more stable.

Some people are already running it across separate terminals

User Louis describes his approach: three computers separately run the doing, checking, and revising, plus a "secretary" specifically for acceptance. He says the biggest benefit after separating them is that the one checking will not make excuses for the one doing the work.

This division of labor sounds crude, but it maps exactly to the multi-Agent collaboration part of the demo. The difference is that what takes 10 minutes to configure in the demo requires first thinking clearly about who reviews whom and with what when actually deploying.

The cost beyond those 10 minutes

User Kai raises a point most people would not think of: building it takes 10 minutes, but maintenance is a long-term cost. The CLAUDE.md or AGENTS.md that the Agent has to read every round will quietly go stale. He scanned 110 active repos with over 3k stars and found that 7 (6%) of the agent docs still pointed to scripts that had been deleted or moved. He suggests adding a path check step in CI.

A few other replies are more down-to-earth. User Jeree says he got stuck for half an hour just configuring the environment, so he will bookmark it first. User Professor An asks: building it in 10 minutes sounds fast, but how well does it actually run? The original post did not answer this question.

Some people are not buying it. User Hale complains that the barrier is so low that even interns can be architects now.

What is valuable is not those 10 minutes

User Jason AI News puts the summary quite well: the key is not building it in 10 minutes, but that the closed loop of execution, checking, and improvement can be reused. A single-Agent demo looks good, but what is truly valuable is the checking mechanism during multi-Agent collaboration.

From the comments, this judgment holds up. The demo proves how fast it can be built; the comments prove how hard it is to write judgment criteria after it is built. Both things are real, but the latter will not appear in a 10-minute video.

If you want to try it, it is recommended to write the checking criteria first before opening the terminal. If you reverse the order, you will probably get stuck at the same place.