Agent governance

Governance is enforcement,
not observation.

Cortega sits between your agents and everything they touch — LLMs, MCP tools, databases, APIs — and enforces the boundary between what an agent is authorized to do and what it actually attempts. Not logged after. Enforced before.

Know who called, what they asked for, and whether they were allowed.

Every request that reaches your agents carries verified identity — JWT, mTLS, or OAuth 2.0. Unauthenticated requests are blocked at the gateway. Virtual keys with full lifecycle management (create, rotate, expire, revoke) are standard — not Enterprise-gated.

Virtual-key lifecycle in the base product

Create, rotate, expire, and revoke keys per team, app, or agent. Full RBAC (viewer, admin, auditor). SSO via Okta or Entra ID. Not paywalled behind an Enterprise tier.

Authorize every inter-agent call

Check inter-agent calls against per-identity access policy before they proceed. Enforce delegation limits.

Zero ambient authority

Agents receive only the permissions their current task requires, for the duration of that task. Least privilege enforced at the gateway — limits the blast radius of any compromised agent.

Two-layer policy on every tool call. Both must pass.

Access control check: is this agent permitted to call this tool? Argument check: do the specific arguments violate your policy, regardless of the agent's stated intent? Both must pass before execution. Neither is probabilistic.

Access control layer

Define which agents, users, and teams can call which tools, models, and MCP servers. Enforced at the gateway on every request.

Intelligent argument check

Even if the agent is authorized, the specific call is inspected. Does it contain PII? Is it trying to access a restricted resource? Is the argument pattern anomalous? Fail-closed on ambiguity.

Approval tracking on consequential actions

When an agent is about to take a consequential action, get a plain-English summary of what it's about to do. Record a documented approval decision, visible in your audit log.

Control what leaves your boundary.

Every LLM call routes through Cortega's egress gateway. You decide which AI providers are approved and which models agents can use. Sensitive data can be detected and redacted before it reaches an external provider.

Approved provider and model lists

Define per-team, per-agent allow/deny lists for providers and models. Block calls to unapproved destinations at the gateway.

Deterministic sensitive-data detection

Local RE2 engine detecting PII and payment-card data (PCI), plus healthcare-adjacent patterns like medical license numbers. No data sent to a third-party scanner. No probabilistic LLM guessing. Ambiguity fails closed.

Output scanning

Guardrails scan model responses too — catching sensitive-data leakage, RAG leakage, and data exfiltration on the return path.

Enforce budgets at the gateway. Not report overspend at month-end.

Team and provider budgets, enforced as hard blocks at the gateway, not after-the-fact alerts. Cortega also detects token velocity anomalies and recursive loop signatures, stopping Denial of Wallet attacks before they drain your API budget.

Team and provider budget enforcement

Set monthly budget limits at the team or provider level. Hard blocks, not alerts.

Rate limiting (RPM/TPM, token-aware)

Enforced per principal. Prevents any single agent or user from consuming disproportionate capacity or triggering provider throttling.

Denial of Wallet detection

Circuit-break on call frequency spikes, token velocity anomalies, and loop signatures per agent session. This is both a cost control and a security control.

Governance that your compliance team will approve.

Deploy Cortega and show your security team the evidence before they sign off on a single agent.

Schedule an Architecture Review →