Runtime Integrations
AgentWarden can receive runtime events from custom agent applications or supported runtime hooks.
Both paths use the same AgentWarden policy model. The difference is where the runtime event comes from and who enforces the returned decision.
Choose an Integration Path
| Customer situation | Recommended path |
|---|---|
| You own the custom agent loop. | Use the AgentWarden SDK. |
| You use a supported coding-agent runtime. | Use the AgentWarden hooks installer. |
| You use a gateway or company AI platform layer. | Emit AgentWarden-compatible runtime events from that layer. |
Supported coding-agent runtimes currently include Cursor, Claude Code, and Codex. Contact the Dynamo AI support team for the current support matrix.
The useful question is: where can AgentWarden observe the action boundary and enforce the decision?
SDK
Use the SDK when your application owns the agent loop.
Your application adds checkpoints before prompts, before tool calls, after tool outputs, and before final responses. AgentWarden returns decisions, and your application applies them.
This path gives the most control because the application can decide exactly where to call AgentWarden and how to handle approval, blocking, replacement, or audit behavior.
For implementation details, see SDK.
Hooks Installer
Use hooks when a supported coding-agent runtime owns the loop.
The hook integration maps runtime events into normalized AgentWarden events, sends them to AgentWarden, and enforces the returned decision inside the runtime.
This path is lower-friction for supported runtimes because the developer does not need to modify an agent application.
For installation shape and runtime behavior, see Hooks Installer.
Gateway or Platform Layer
Some teams already route agent traffic through a company platform or gateway.
AgentWarden can fit there if that layer can emit the right runtime events: prompt, pre-tool, post-tool, and final-response boundaries. This can be useful when a team wants centralized coverage, but it only protects the boundaries that the platform can actually observe.
Full coverage may require more than one layer when prompts, tool calls, local tools, remote tools, and coding-agent runtimes happen in different places.