Connecting an agent
Kaizen attaches to your agent in one of five ways. The deeper you attach, the harder it is for an agent to route around you, and the more Kaizen trusts what it sees. Start anywhere and deepen later.
| How you attach | Trust | What it sees |
|---|---|---|
| SDK inspect | Cooperative | Each tool call you wrap in code |
| Framework adapter | Cooperative | Tool calls via LangChain, CrewAI, LlamaIndex, Pydantic AI, Semantic Kernel, Vercel AI |
| MCP shim | Chokepoint | Every tool call through one MCP server |
| Egress proxy | Ground truth | The real outbound connections the agent makes |
| Sandbox / eBPF | Ground truth | Kernel-level, next to a self-hosted sandbox |
Open Connect agent in the console for copy-paste setup for each. See also Observation depth.
Observation depth per agent
Each agent shows the deepest way it has been observed and its trust tier. An agent attached only by a cooperative hook gets a prompt to deepen it to ground truth, and every verdict is tagged with the feed it came from.