Skip to main content

Govern People and Their Agents

You answer three questions. Who is this person? What can they reach? What can their agent do with that access? Infrastream answers all three in manifests. Each change to an answer is a pull request.

In short
  • A person belongs to groups. A group gets a tier on a project: read, dev or admin.
  • An agent that signs in as a person gets the tiers of that person, and no more.
  • A deployed agent reaches only what its accessControl block lists.
  • A release stage can require approval from named groups.
  • No tier can read a secret payload, change IAM or impersonate a service account.

Six controls​

Each control answers one question. You set each one in a place that your team can read and review.

ControlThe questionWhere you set it
GroupsWho belongs to which team?An OrganizationUserGroup manifest
TiersWhat can this team do in this project?The permissions block of a Project, Environment or OrganizationalUnit: administrators, contributors, viewers
Release approvalWho must say yes before a version moves on?The approvalPolicy of a stage in a ReleaseTrack
Branch rulesWho can merge a pull request?The branch rules of your GitHub repository
Agent access boundaryWhat can a deployed agent reach?The accessControl block of an Agent manifest
Withheld permissionsWhat can nobody do?Built in. See What no tier can do.

What each kind of agent inherits​

AgentWhat it inheritsStatus
BeckThe engine applies what a person merges. Your branch rules decide how strictly your team reviews it.Available
The AI tool of an employee, through MCPThe tiers of the person who signs in. The consent screen can only narrow the access.Pilot
An agent that your company deploysIts own identity, and the resources that its accessControl block lists. It cannot create IAM bindings outside its project or change egress rules.Available. The agent layer is early.

Recipes​

Give a new employee the access of their team​

  1. Add the person to the group of their team. List the OrganizationUser name, not an email address.

    apiVersion: lowops.manifests.v1
    kind: OrganizationUserGroup
    metadata:
    name: payments-contributors
    organization: acme-corp
    spec:
    description: Developers with contribution access to AstraPay production.
    members:
    users:
    - tech-lead
    - jane-smith # added
    external: []
  2. Make sure that the permissions block of the project lists the group.

  3. Open a pull request and merge it.

You can ask Beck to open this pull request. Say, "Add jane-smith to the payments-contributors group."

Give a team more access in test than in production​

A project is in one environment. So you can give the same team a different tier in each environment. The payments team gets the dev tier in integration and the read tier in production.

environment/integration/project/payments/payments.yaml
spec:
permissions:
contributors:
groups: [payments-team]
environment/production/project/payments/payments.yaml
spec:
permissions:
viewers:
groups: [payments-team]

The agents of the team now get the dev tier in integration and the read tier in production.

Require approval before production​

Add an approvalPolicy to the production stage of the ReleaseTrack. The release waits in the portal until a member of the group approves.

spec:
releaseStages:
- environments:
- prod
approvalPolicy:
stakeholder:
groupEmail: retail-cloudops-leads

You can combine approvers with anyOf, allOf or quorum. See the ReleaseTrack reference.

A stage with no approval policy needs no approval

Add a policy to each stage that must wait for a person. Do not assume that every stage waits.

Remove access when a person leaves​

  1. Delete the person from each group that gives access.
  2. Open a pull request and merge it.
  3. Check the project in the Google Cloud console. Ask the person to open the resources that their tier allowed. The access must be gone.

Check your branch rules​

Beck proposes a change as a pull request. Your branch rules decide whether a review must happen before the merge. The GitHub app of the assistant and the organization admins can bypass the rules. Check this setting. Do not assume it.

When a change takes effect​

  • A new group member gets access when the engine applies the merged manifest.
  • A tier change reaches an MCP token at the next sign-in, not at a refresh. An access token that is already out stays valid until it ends.
  • The gcp MCP server examines the tier of the project on each call. A project tier change takes effect up to 90 seconds after the portal reads the applied manifests.

What these controls do not cover yet​

  • The MCP is in pilot. Customers cannot use it yet.
  • Production stays with people. In integration, Infrastream refuses a proven agent that starts a release. It also refuses a proven agent that starts, approves, rejects, retries or rolls out a stage after integration. See Agent Access Model.
  • A proven agent is a caller with a signed assertion, or an IAP service account. Every MCP client counts as an agent, so use the portal for these actions. The MCP production route does not run yet.
  • A cap for one agent is not open. We plan it, and we defer it for now.
  • AI spending limits (coming soon). You will set one daily and monthly limit for the company and a different limit for any person. The limits count use of the Infrastream assistant only, not other AI tools. Each employee will start with $15 a month. Members of the ai-admins group are exempt. They can change or remove the default.
  • The record is the pull request, the plan and the run history. It is not a complete audit of every action.

Next​