𝗧𝗵𝗲 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜 𝗚𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 𝗙𝗿𝗮𝗺𝗲𝘄𝗼𝗿𝗸
Static PDF policies cannot govern dynamic AI agents. If you require manual approval for every agent action, you will lose speed.
To scale, you must move from Human-in-the-Loop (HITL) to Human-on-the-Loop (HOTL). In HITL, humans are bottlenecks. In HOTL, humans act as architects who build guardrails and monitor systems.
The Autonomy Spectrum
You should not treat a procurement agent the same as a DevOps agent. Categorize your agents by their decision-making authority:
• Advisory (Low Autonomy): The agent suggests actions. The human executes them. • Semi-Autonomous (Medium Autonomy): The agent acts within a safe zone. It asks for help only when it hits a limit. • Fully Autonomous (High Autonomy): The agent manages goals from start to finish within a sandbox.
Example: A support agent can issue refunds up to $50. If a request is $51, the agent must stop and ask a human. This prevents authority drift.
Stop Relying on Prompts for Security
System prompts are suggestions, not rules. They are probabilistic. In production, a suggestion is a vulnerability. You must separate reasoning from enforcement.
Use a layered defense:
- Prompt Layer: Provides intent and guidelines.
- Guardrail Layer: A deterministic middleware that validates actions against hard rules.
- API Layer: Enforces identity and access management at the resource level.
This setup prevents prompt injection. If the guardrail is at the middleware level, a trick in the prompt will not work.
Observability and the Kill Switch
Standard logs are not enough. You need to log the Chain-of-Thought (CoT). You must know why an agent thought an action was correct.
To prevent cascading failures in multi-agent workflows, implement a Kill Switch Protocol. It must:
- Revoke all active tokens.
- Terminate all execution threads.
- Freeze the state for audit.
- Notify an engineer with the last five reasoning steps.
Dynamic Permissioning
Agents should have zero standing privileges. Use Just-in-Time (JIT) access. An agent requests a short-lived token only when it needs to call a specific API. This ensures the agent identity remains tied to human-approved intent.
Key Takeaways for Platform Engineers:
- Trasladar la gobernanza de los prompts al middleware.
- Implementar acceso JIT para prevenir brechas de seguridad.
- Registrar los pasos de razonamiento, no solo las entradas y salidas.
- Usar límites estrictos de tokens para evitar fallos por bucles recursivos.
Fuente: https://dev.to/omnithium/the-agentic-ai-governance-framework-balancing-autonomy-and-control-2occ
Comunidad de aprendizaje opcional: https://t.me/GyaanSetuAi