- In a Google DeepMind demo, an agent was a main instruction file plus a few skill files. Moving it to a back end meant copying that file.
- Our reading: when an agent is files, the human work is writing clear instructions and judging the output. The talk did not say how to train for that.
Philipp Schmid of Google DeepMind showed an agent built from a handful of text files. He did so in a talk on the AI Engineer show. His demo points to a shift in craft. Less of the work is writing code. More of it is writing clear instructions and checking what the agent produced.
What was said
Schmid presented Google's Gemini Interactions API and a managed cloud sandbox. That is a remote computer an agent gets for itself. He described how use has changed. People once sent a prompt and read a reply. Now, a speaker said, you give the model a goal and tell it how to verify the goal and how to deploy it. Then you hope it runs for hours or even days.
Then came the demo, a podcast-style app. When the talk opened up its agent, a speaker on the show said: "And our agent is just files."
One file, called AGENTS.md, loads when the environment starts. It becomes part of the system instruction, the standing brief the model reads first. It tells the agent to research, write scripts for different people, generate music and speech, and mix everything together. Skills do the rest. One skill explains how to run speech generation. It comes with a short Python file that calls other Gemini models.
To move the agent from the interface to a back end, the speaker said you copy that file and create the agent. An agent can also build its own environment, a speaker on the show said. That works "without writing any Terraform, without thinking about Kubernetes, microservices, Firecracker, or anything else."
Why it matters
Our reading: if an agent is a brief plus a few skills, then the brief is the product. Whoever writes it decides what the agent does, in what order, and what counts as done.
The talk's own framing supports this. Telling the model how to verify its goal means a person must first define what a good result looks like. After a long run, someone still has to judge what came back.
For managers, the people who know the work well may matter more than the people who write the most code. The speaker also described teams sharing agents, so one team's instructions could become another team's tool. That makes the quality of those files a shared concern.
The other side
This was a product demo from the company that sells the platform. The example was a podcast-style app, where a weak result costs little. Nothing in the excerpts tested an agent on high-stakes work.
The talk also did not say how people learn to write good instructions or to judge long runs. Nor did it say who is accountable when an agent that ran for days gets it wrong.
Code has not vanished either. The speech skill in the demo still relied on a Python file. The shift is one of emphasis, not a clean break.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Why AI Agents Should Have Their Own Sandbox — Philipp Schmid, Google DeepMind (2026-10-07)
