Google's most important agent announcement is not the agent. It is the file. In a May 19 developer post, Google said Gemini API users can define custom managed agents with markdown files such as AGENTS.md and SKILL.md, then run those agents through the Antigravity agent inside isolated Linux sandboxes. [1] Wednesday's paper argued that Google was turning agent instructions into platform files. Thursday's follow-up is sharper: Google is not merely recognizing a convention from coding-agent users. It is trying to make that convention deployable infrastructure.

That distinction sounds small until one remembers how platforms win. They do not always win by having the cleverest demo. They win by deciding where developers put the thing that matters. The browser made the URL a civic object. Git made the commit a social object. Docker made the container file an operational object. Google's Managed Agents post is asking for a similar promotion: the agent's instructions, skills and data should be versionable files that can be registered with a cloud service rather than private text pasted into a chat window. [1]

The source stack supplies two different Google voices. The main Innovation & AI page treats the week as a catalog: I/O 2026, Gemini, developer tools, infrastructure, provenance and product announcements arranged as a front door to the company narrative. [2] Sundar Pichai's I/O transcript gives the grander thesis, calling this the "agentic Gemini era" and placing agents inside a vast full-stack story of custom silicon, models, products, consumer assistants and enterprise demand. [3] The Managed Agents post is less theatrical. It is the shop-floor receipt.

Managed Agents, Google says, let a developer spin up an agent with a single call. The agent can reason, use tools, execute code and manage files inside an isolated, ephemeral Linux environment. It can browse the web, fetch live data and preserve session state across follow-up calls when it receives or creates an environment. [1] This is not a better autocomplete. It is a claim about agency as hosted runtime.

The post's nouns matter. A "harness" is not a model. A "remote Linux environment" is not a chat transcript. A "secure cloud sandbox" is not a prompt. When Google says it is opening up the agent harness and infrastructure behind Deep Research so developers can build their own custom managed agents, it is selling a place for work to happen. [1] In the chatbot phase, the user brought the task to the model. In this phase, the platform provisions a little workplace for the model and then asks the user to define the worker.

That worker is defined in files. Google says developers can extend the Antigravity agent with their own instructions and skills, and define everything in markdown files like AGENTS.md and SKILL.md rather than writing complex orchestration code. [1] This is the quiet politics of developer tooling. A team that adopts those names is not just choosing a syntax. It is teaching its repository how to describe work to an agent in a way another tool might later understand.

The mainstream version of this story will understandably focus on I/O scale. Pichai's transcript is full of immense numbers: more than 3.2 quadrillion monthly tokens across Google's surfaces, more than 8.5 million developers building monthly with Google's models, roughly 19 billion API tokens per minute, and $180 billion to $190 billion in expected annual capex after a 2022 figure of $31 billion. [3] Those numbers are not decorative. They explain why a hosted agent runtime is a business priority. When compute becomes that expensive, the interface to compute becomes strategic.

But the more consequential standard may arrive in the humble form of a markdown file. Developers already keep READMEs, test commands, lint rules, deployment notes and architectural decisions in repositories because the repository is the common memory between humans and machines. If agent instructions stay trapped inside product settings, every tool becomes an island. If they become files, the behavior can be reviewed, diffed, copied, audited and forked.

That is why this announcement belongs in a newspaper rather than a product roundup. The agent market is often covered as a model contest: whose coding score is higher, whose demo is faster, whose assistant can book a trip with fewer clicks. Google's move points to the layer beneath the contest. The question is not only which model acts. It is who owns the contract that tells acting software what it may do.

There is an institutional meaning here, too. Versioned files have a discipline that conversational interfaces lack. A manager can review a change to AGENTS.md; a security team can ask why a skill was added; a lawyer can discover the instructions that governed a sensitive workflow; an engineer can bisect a regression in behavior. A prompt in a proprietary UI may govern real work, but it is hard to govern in return. A file can be blamed.

Google's announcement also collapses two ambitions that companies usually keep apart. One is the consumer ambition described in Pichai's I/O remarks, where Gemini Spark becomes a 24/7 personal agent running on dedicated virtual machines and acting across tools under user direction. [3] The other is the developer ambition in the Managed Agents post, where the same broad idea becomes an API, a harness, a sandbox and a file structure. [1] The public sees the assistant. The builder sees the runtime.

The divergence with X is predictable. Social media is built to reward visible capability: the clip of an agent browsing, coding, fixing, booking or failing spectacularly. It is less good at noticing directory formats. Mainstream product coverage, meanwhile, often follows the keynote's hierarchy and treats the developer post as one more item in a long I/O list. The useful story sits between those habits. A platform is forming around agent work, and the file names are part of the land grab.

The risk is that a versioned instruction file becomes another standard-shaped moat. If Google succeeds, it will have made agent behavior more portable in one sense and more Google-shaped in another. The company can say, fairly, that developers need less scaffolding and sandbox management. [1] It can also benefit when the conventions developers adopt point naturally back toward Gemini API, Antigravity and Google AI Studio.

That is not sinister. It is how developer ecosystems are built. The best platforms make useful defaults until the defaults become assumptions. The caution is that assumptions about agent authority, tool access and skill loading will soon matter outside software teams. An autonomous agent that can browse, execute code and manage files is not merely a helper. It is a participant in an operational environment. The documents that instruct it deserve the same seriousness as deployment manifests or access policies.

Pichai's I/O transcript repeatedly frames AI through scale: tokens, users, products, capex, silicon, speed and savings. [3] The Managed Agents post translates that scale into a developer primitive. [1] That translation is the story. The age of the agent will not be settled by a single model release. It will be settled by boring things: where instructions live, how sandboxes are provisioned, how state persists, how tools are called, how changes are reviewed.

For readers, the practical test is simple. Watch whether AGENTS.md and SKILL.md remain examples in one Google blog post or become artifacts other tools feel obliged to parse. Watch whether enterprises treat agent instructions as compliance documents. Watch whether open-source projects begin accepting pull requests to agent files the way they accept changes to test scripts. If that happens, Google will have done more than ship another agent service. It will have helped turn agent behavior into code-adjacent infrastructure.

The word "agentic" has already been worn thin by marketing. The file has not. A file can be diffed. A file can be owned. A file can move through review. A file can sit in a repo long after the demo has gone stale. Google's announcement matters because it points agentic work back toward the oldest discipline of software: put the important thing where everyone can inspect it.

-- ANNA WEBER, Berlin