Skip to content

Core Concepts

AgentSpine is designed as a decoupled control plane for AI agents. Unlike traditional agent frameworks, AgentSpine does not dictate how you build your agent (it works with LangChain, CrewAI, Autogen, or custom code); instead, it manages the governance, safety, and reliability of the agent’s actions.

In a production environment, an agent’s lifecycle follows this flow when integrated with AgentSpine:

  1. Intention: The agent decides to perform an action (e.g., “Send an email”).
  2. Consultation: The agent calls the AgentSpine SDK to validate the action.
  3. Orchestration: AgentSpine runs the action through the Policy Pipeline.
  4. Enforcement: Policies (Deduplication, Risk Scoring, Guardrails) approve or block the action.
  5. Execution: If approved, the action is dispatched to an executor.
  6. Persistence: The result is logged to the Event Timeline for auditing and self-improvement.

The heart of AgentSpine. It evaluates every requested action against a set of rules. Policies can be:

  • Synchronous: Immediate pass/fail based on hard rules.
  • Asynchronous: Triggering external approvals or complex analysis.

Prevents agents from repeating the same action unintentionally (e.g., sending the same email twice).

  • Hard Dedupe: Exact match on action and parameters (powered by Redis).
  • Semantic Dedupe: (Coming Soon) Detects conceptually identical actions.

AgentSpine provides a generic boundary for executing actions. This allows you to swap out implementations (e.g., from a local function to a remote webhook) without changing the agent’s core logic.

A chronological record of every decision, action, and result. This timeline is the foundation for:

  • Audit Logs: Why did the agent do this?
  • Replay: Debugging complex agent failures.
  • Reward Logging: Providing feedback to the agent’s underlying model for training.

AgentSpine supports two primary integration patterns:

The SDK runs inside your Python process and connects directly to your databases. This provides the lowest latency and is ideal for Python-based agent stacks.

A standalone FastAPI server that exposes an HTTP API. This allows non-Python agents to interact with the control plane and enables the centralized Dashboard.