SOIT Docs
Core

Govern

Enforce SOIT permissions, secrets, egress, audit, versioning, and policy boundaries.

Govern is the control plane that keeps agent execution bounded. SOIT assumes agents will touch sensitive systems, so identity, workspace scope, secret handling, egress, audit, and capability allowlists are platform concerns — not application-level afterthoughts.

What Govern enforces

  • Tenant and workspace scoping on resources and runtime state.
  • RBAC and resource-level permission checks.
  • Vault-backed secret references instead of plaintext secrets.
  • Deny-by-default egress policy for external calls.
  • Policy-aware ports for models, tools, vector search, storage, and secrets.
  • Immutable versions and release pointers for agents, workflows, and plugins.
  • Audit records for privileged operations and governed runtime behavior.
  • Per-key limits, service principals, and content-free runs for callers of the OpenAI-compatible gateway and the MCP server.
  • Budgets that alert at thresholds and refuse calls at a hard stop.

Where enforcement happens

Governance is applied at three layers, and none of them can be skipped by application code:

  1. Request boundary — every request resolves a RequestContext (tenant, user, workspace, role) before any business service runs. Business APIs never accept tenant or workspace ids as free parameters.
  2. Data boundary — scope-aware repositories add tenant/workspace filters to every workspace query, and unique constraints include the scope columns.
  3. External boundary — every call to a model, tool, vector store, object store, or secret backend crosses a kernel port that applies timeouts, retries, rate limits, audit logging, secret resolution, and egress checks consistently.

Secure-by-default posture

  • Egress is deny-by-default; external destinations must be explicitly allowed.
  • Secrets are references only — a secret:name pointer resolved through Vault at the gateway, never a plaintext value in business rows, specs, or logs.
  • High-risk runtimes stay disabled unless explicitly enabled.
  • Publish gates can require passing regression evaluation before an agent version goes live.

Seeing governance work

The demo workspace includes a 20-minute governance tour covering permissions, secret inventory, call audit, cost attribution, replay, and regression — each step backed by real runtime evidence rather than configuration screenshots.

Design principle

Business code should not bypass governance. External systems are accessed through kernel ports and adapters so timeouts, retries, audit, rate limits, secret resolution, and egress checks are applied consistently — the import rules make bypassing them a build failure, not a code-review debate.

Observe

Inspect SOIT runtime behavior through runs, traces, costs, approvals, feedback, and replay evidence.

Capabilities

The concrete SOIT product modules that make up the governed agent runtime.

On this page

What Govern enforcesWhere enforcement happensSecure-by-default postureSeeing governance workDesign principleRelated capabilities