Agent Workflow
A model that detects something is only useful if that detection triggers the right response. Agent Workflow is the layer between what the model sees and what the system does.
The Problem It Solves
Edge AI inference produces structured outputs — class labels, confidence scores, bounding boxes. Without an automation layer, those outputs sit on the device or get logged somewhere. Acting on them — sending an alert, calling an API, triggering an actuator — traditionally requires custom firmware development or external integration work.
Agent Workflow eliminates that gap. It provides an LLM-powered decision layer that interprets inference outputs and executes configured responses, without requiring code changes or firmware modifications.
How the Workflow Operates
LLM-CJSON-OPENCLAW pipeline:
- Inference output — The edge model produces structured results (CJSON format), describing what was detected with confidence and context
- LLM decision — A language model interprets the structured output against configured rules and context, determining what action is appropriate
- OPENCLAW execution — The action layer executes the response: calling an external API, sending a notification, writing to a database, or triggering a physical output
This pipeline runs on-device or at the edge. It does not require a round-trip to a cloud service for every inference event.
What Agent Workflow Replaces
| Approach | Limitation | Agent Workflow Alternative |
|---|---|---|
| Blockly visual programming | Limited expressiveness, no LLM reasoning | Natural language rules, LLM decision logic |
| n8n / Dify workflow tools | Separate infrastructure, manual integration | Built into the deployment, no separate service |
| Custom firmware logic | Requires engineering resources per use case | Configured through the platform interface |
| Manual monitoring | Reactive, requires human attention | Automated response to defined conditions |
Example Responses
The OPENCLAW layer can execute a range of configured responses without custom code:
- Messaging — Send an alert to Slack, WeChat Work, or any webhook endpoint when a detection threshold is reached
- API calls — Trigger an external system (ERP, SCADA, dashboard) when a specific condition is detected
- Logging — Write structured detection records to a database or time-series store
- Event routing — Fan detection events out to multiple downstream systems simultaneously
- Conditional logic — Different responses for different detection outcomes (e.g., "defect type A → alert; defect type B → log only")
Integration Philosophy
Agent Workflow connects to systems your team already uses through standard interfaces. It is not a replacement for your existing tooling — it is the bridge that routes edge AI events into the workflows and dashboards you already operate.
The inference runs at the edge. The decision runs at the edge. Only the final action output — a webhook call, a database write — leaves the local environment.