Aegis Runtime
Your agents can act. We define how far.
The AI Action Envelope™ gives every autonomous agent a bounded, enforced authority — checked at the moment it acts, not just described in a policy document.
A dashboard tells you what an agent did. Aegis Runtime decides whether it's allowed to happen at all — before the action reaches your systems.
We don't replace your agent. We sit between it and what it's allowed to touch.
Every action request passes through the gateway before it reaches a business system — evaluated against the agent's registered authority in real time.
Governance applies to specific actions, not vague access levels.
Each tool an agent can call has a stable identity — refund.issue, customer.lookup, ledger.delete — so authority can be granted and denied precisely.
Customer Refund Agent — permitted tools
Denied by default
- allowed actions
- Check refund eligibility · Issue customer refund
- restrictions
-
- Maximum autonomous refund: $400
- Maximum refund requiring approval: $5,000
- Refunds above $5,000: deny
- Approved customer accounts only, no bulk refunds
- Production access only
- validity
- 1 Sept 2026 – 30 Sept 2026
- owner / approver
- Customer Operations / Finance Governance
From first registration to production enforcement.
Every agent moves through the same guided path — nothing reaches production until its envelope has been reviewed.
Set up your organisation
Create the workspace and invite the administrators who'll own registration and approvals.
Register the agent
Name it, assign a business and technical owner, and state its purpose and environment.
Assess its risk
Aegis Risk runs a guided questionnaire — financial impact, data sensitivity, reversibility, blast radius — and produces a risk profile.
Define the envelope
Set permitted actions, financial limits, approval thresholds and validity dates.
Simulate before enforcing
Run the envelope against real traffic first: "this would have blocked 17 actions, and required approval for 6."
Activate and monitor
Turn on production enforcement, route approvals, and revisit the risk rating as the agent's role changes.
One place to see every agent's authority — and every decision it made with it.
| Agent | Risk | Status |
|---|---|---|
| Customer Refund Agent | HIGH | Active |
| HR Information Assistant | MEDIUM | Active |
| Internal Knowledge Agent | LOW | Active |
| Supplier Payment Agent | CRITICAL | Suspended |
- Agent registration
- Risk assessments
- Envelope creation & changes
- Policy decisions
- Approved actions
- Denied actions
- Suspensions
- Human approvals
- Actual executions
- Failed executions
- Authority violations
Three ways in, depending on how the agent is built.
Python SDK
pip install governance-kernel
from governance_kernel import GovernedGateway
gateway = GovernedGateway(
agent_id="finance-refund-agent",
api_key="client-issued-key"
)
result = gateway.execute(
action="refund.issue",
arguments={
"customer_id": "customer-123",
"amount": 4000
}
)
Hosted gateway
POST /v1/agents/finance-refund-agent/actions
{
"action": "refund.issue",
"amount": 4000,
"customer_id": "customer-123"
}
← {
"decision": "REQUIRE_APPROVAL",
"request_id": "req-1042",
"reason": "Amount exceeds
autonomous limit"
}
Governed tool endpoint
// The agent sees an ordinary // MCP tool. // // Every invocation is routed // through the enforcement // layer first — no change // to the agent's own // reasoning loop.
A base platform fee, plus the scale of what you enforce.
Priced on governed AI actions evaluated by the gateway, not on the number of agents you register — so consolidating agents to save money is never the incentive.
- Up to 5 registered agents
- Development & staging environments
- Basic Action Envelopes
- Standard audit history
- Email support
- More agents & environments
- Production enforcement
- Risk Registry & human approvals
- SSO and role-based access
- Extended audit retention
- High-volume enforcement
- Private or dedicated deployment
- Advanced identity integration
- Custom policy & integrations
- Security reviews & SLAs
These ranges are illustrative starting points, not confirmed pricing — final packages are scoped against your agent count, environments and approval workload.
Put bounded authority around your first AI agent.
Aegis Runtime is currently onboarding selected design partners. Work with us to define an AI Action Envelope™, integrate your agent, simulate its authority boundaries, and move toward controlled enforcement.