Safety Kernel
A safety kernel is the policy decision point of an agent control plane — a dedicated service that evaluates every agent action before dispatch and returns a decision such as ALLOW, DENY, REQUIRE_APPROVAL, or ALLOW_WITH_CONSTRAINTS.
Definition
A safety kernel is the policy decision point of an agent control plane — a dedicated service that evaluates every agent action before dispatch and returns a decision such as ALLOW, DENY, REQUIRE_APPROVAL, or ALLOW_WITH_CONSTRAINTS.
How it works
In Cordum, the Safety Kernel is a gRPC decision engine the scheduler calls before dispatching any job. It evaluates the action against versioned policy rules — matching on tenant, topic, capabilities, risk tags, actor identity, and MCP tool context — and returns a normalized decision. Because the scheduler treats the kernel as part of the hot path, it uses short client timeouts and a circuit breaker to protect throughput, with an optional decision cache to avoid re-evaluating identical requests.
The decision set
Decisions are normalized to a small, deterministic set: ALLOW lets the job proceed; DENY blocks it and returns a reason to the caller; REQUIRE_APPROVAL holds the job in a human approval queue; and ALLOW_WITH_CONSTRAINTS lets it proceed with scope, rate, or redaction limits attached. A throttle decision is also available when a velocity rule's rate budget is exhausted. Output policy is evaluated separately, after execution, and can allow, redact, or quarantine a result before it returns to the caller.
Frequently asked questions
How many decision types does the Safety Kernel return?
The kernel normalizes decisions to a small set. The four core decisions are ALLOW, DENY, REQUIRE_APPROVAL, and ALLOW_WITH_CONSTRAINTS, with an additional THROTTLE decision when a velocity rule's rate budget is exhausted.
Does the Safety Kernel slow down agents?
It is designed for the hot path. The scheduler calls it with short timeouts and a circuit breaker, and an optional decision cache lets repeated identical checks return without re-evaluation, keeping per-action overhead low.
Related reading
Govern your AI agents with Cordum
Cordum is the agent control plane: policy-before-dispatch enforcement, human approvals, and a tamper-evident audit trail for autonomous AI agents.