Agent OS · Policy-Compiled

Every agent in your org, in one control plane

Register, scope, and gate every AI agent against compiled policy. If it isn't in Agent OS, it isn't running.

agent-os · registry
312Agents
100%Policy coverage
9Teams
support-agent · support-agents governed
billing-agent · finance-strict governed
intake-agent · hipaa-vpc governed
sandbox-bot · quarantined

Governing agents at

KavanoNordvikHalcyonBroadfieldVeratech

A registry, a policy engine, and an enforcement layer

Central registry

Every agent, its owner, its scopes, and its policy discoverable in one place across clouds and vendors.

Policy engine

Compile guardrails from code and bind them to identity, model, tools, and data at every call.

Real-time enforcement

Decisions happen inline a blocked tool or model never reaches your users, and every call is logged.

Identity-aware

Scopes bound to who, not just what

pactflow resolves the caller's identity through your IdP, then grants each agent only the permissions that identity and policy allow. Least privilege, enforced per request.

How enforcement works
grant agent="support" for="role:agent.cs":
    tools:  allow("zendesk.read", "refund.lt(50)")
    data:   scope("tenant")
    deny:   "export.*"
✓ resolved via Okta · enforced inline

Deploy anywhere

Your cloud, your VPC, your rules

Run Agent OS as managed SaaS or install it entirely inside your own infrastructure. Models can be pinned so data never leaves your boundary. SSO, SCIM, and RBAC are built in.

Compare deployment options
deployment
region: vpc-us-east isolated
SSO · SCIM on
models pinned in-boundary on
audit → Splunk streaming

One lifecycle, from register to retire

1

Register

Every agent declares an owner, a purpose, and requested scopes at birth. Unregistered traffic is quarantined on sight, like sandbox-bot above.

2

Scope

Compiled policy binds identity to model, tools, and data. The agent gets exactly what its owner's role and its stated purpose justify.

3

Operate

Enforcement runs inline on every call, not in a nightly review. A denied tool never executes and a denied export never leaves.

4

Retire

One action revokes credentials, drains in-flight work, and archives the audit trail. No orphaned keys, no zombie agents.

agent-os · lifecycle
pricing-agent · no registration found quarantined
pricing-agent · owner + scopes declared in review
pricing-agent · finance-strict bound governed

Model governance

The model an agent runs is a policy decision

Each team routes through an allowlist of approved models, pinned to a version. An agent never silently upgrades: a model change is diffed and routed through Change Control like any other change. Fallback routes carry the same policy, and you bring your own model keys, so pactflow never proxies your model billing.

How model changes are reviewed
route team="finance":
    models:   allow("anthropic:claude-sonnet@2026-02")
    fallback: "openai:gpt-5@2026-05"  # same policy
    keys:     byo("kms:finance-llm")
✓ pinned · upgrades go through Change Control

The audit log

Every decision is a signed record

Allow, block, or quarantine, each decision lands in a hash-chained log: who called, which identity resolved, which policy version fired, and what was permitted or denied. Query by agent, team, policy, or time window. Stream to Splunk or Datadog, or export the compliance mapping your auditors actually asked for.

Audit log reference
audit log
support-agent · zendesk.read · support-agents v7 allow
billing-agent · export.customers · finance-strict v11 block
sandbox-bot · unregistered call quarantine
chain verified · stream → Splunk signed

Put every agent under one roof.