Skip to main content

Infrastream for Employees and Their Agents

Your employees already use AI to build tools and to fix problems. Infrastream gives that work a safe place to land. The employee asks. The agent proposes. A person promotes.

In short
  • Employees describe a change in plain words. They do not need the Google Cloud console.
  • Agents work through pull requests. Each pull request shows a plan of what will change.
  • Your company decides who can reach what. The rules are files that your team can read.
  • A person merges the change. A release stage can require approval from named groups.

The problem​

An AI-built tool can start on one person's laptop or in one person's account. Then nobody else can see it, secure it or pay for it. When that person leaves, the tool can stop, and nobody knows who owns it.

Infrastream gives each tool a home that the company owns. The tool runs on Google Cloud, behind your company sign-in or behind the sign-in page of your clients. Apps scale to zero when nobody uses them. A fixed base stays on: the gateway, the network, the server of the assistant and any database.

Three kinds of agent​

Infrastream works with three kinds of agent. They differ in who uses them and in the boundary around each one.

AgentWho uses itIts boundaryStatus
Beck, the Infrastream assistantMembers of your own organization. Not your clients.Beck does not call cloud APIs to change your infrastructure. It opens a pull request. A person merges it.Available
The AI tool of an employee, through the Model Context Protocol (MCP)The employee who signs in with the company sign-inThe agent gets the tiers of that person, and no more. Its tool to propose a change writes a pull request on a feature branch.Pilot. Customers cannot use it yet. It is rolling out to production.
An agent that your company deploysYour apps and your staffIt has an identity of its own. It reaches only the databases, secrets and topics that its accessControl block lists.Available. The agent layer is early.

Read The Agent Path for the MCP. Read Agentic Governance for deployed agents.

Why agents do well here​

  1. Your infrastructure is files. An agent reads and writes short, typed manifests. It does not click through console screens. Your team reads the same files.
  2. Each change has a plan. You cannot read every line that an agent writes. You can read the plan.
  3. The rules stay when the model changes. Beck uses the models that your company enables. A new model does not change the rules.
  4. An app and an agent start with nothing. Each app has its own identity, and that identity starts with no permissions. A deployed agent reaches only what its manifest declares.
  5. Non-production is the place to try. Agents change non-production. The release track moves a change on, and a stage can require approval.

What your company keeps​

The manifests and their history stay in the repository of your company. Git holds each change. The state database holds each run.

A pull request shows who asked for a change and who merged it. A release stage that has an approval policy also records the approvals. See Core Concepts.

What this is not

This record is not a complete audit of every action. The Google Cloud audit log of the MCP reader names a service account, not the person. Audit for each person in Google is a later phase.

What each person does​

PersonWhat they doWhere to start
An employee who is not technicalAsks Beck for a change, reads the summary and tells an approver.Start Here If You Are Not Technical
A team lead or an ownerSets who can reach what, and which group can approve a release stage.Govern People and Their Agents
A developer or an engineerConnects an AI agent and tests a change in non-production.The Agent Path

Is it right for you?​

Infrastream is for teams with more than one app or more than one environment. For one small server with no sign-in and nothing to deploy, Infrastream is more than you need.