Govern
How to Govern Semantic Kernel Agents
Semantic Kernel lets you build agents with plugins and planners. Cordum governs which plugins execute, when, and under what constraints.
The problem with ungoverned Semantic Kernel agents
- Plugins expose functions to the AI. The planner decides which to call. No policy layer between decision and execution.
- Filters provide hooks but building governance inside filters couples policy to application code.
- Multi-step plans execute sequentially. If step 3 is high-risk, you only find out when it runs.
- Enterprise teams across multiple services end up with inconsistent governance.
How Cordum governs Semantic Kernel
1
Route Semantic Kernel tasks through Cordum
Submit kernel execution requests as Cordum jobs. Safety Kernel evaluates requested plugins.
const job = {
type: "semantic_kernel.agent.run",
payload: {
agent: "customer-service-agent",
plugins: ["OrderPlugin", "EmailPlugin", "CRMPlugin"],
planType: "stepwise",
goal: customerRequest,
},
};
await cordumClient.jobs.submit(job);2
Centralize plugin access policies
One policy bundle governs all Semantic Kernel services. No per-service filter implementations.
3
Pre-plan governance
Evaluate the full plan before execution. Flag high-risk steps or require approval for sensitive plugins.
4
Consistent audit across services
Centralized audit trail regardless of how many services use the kernel.
Semantic Kernel native vs Cordum governance
| Area | Semantic Kernel Native | With Cordum |
|---|---|---|
| Plugin access governance | Filters (code-coupled) | Declarative policy rules, centralized |
| Pre-plan evaluation | Planner executes without policy check | Evaluate full plan before execution |
| Approval workflows | Must build custom in filters | Built-in with policy-linked approvers |
| Cross-service consistency | Per-service filter implementations | One policy bundle across all services |
| Audit trail | Telemetry/logging (per-service) | Centralized structured run timeline |
FAQ
Does Cordum replace Semantic Kernel filters?
It can. You can still use filters for non-governance concerns like logging and telemetry.
Does this work with both C# and Python SDKs?
Yes. Cordum governs via CAP v2, which is language-agnostic.
Related guides
Ready to govern your Semantic Kernel agents?
Start with the open-source Cordum platform. Add policies, approvals, and audit trails in minutes.