Agent Control
Workflow
Workflow combines LLM calls, tool calls, conditions, agent steps and human approval in a single flow. A flow starts manually, on a schedule, from a webhook or through the API.
Scope
Flows are defined in the panel. There are five step types: LLM call, tool call, condition, agent and human approval. A flow has at most 20 steps.
- Directed flow graph and conditional branching
- Interruption-tolerant execution and checkpoints
- Resume from where it stopped
- Long-running flows and versioning
- Step-by-step run history
- Token and cost tracking at every step
Why fixed-step flows
In enterprise systems, the core structure is designed as a deterministic workflow. The flow spells out which step runs when, which path to take, and where human approval is required. Agent behavior is added where needed, within the limits of the Policy Engine.
- Traceability and easy auditing
- Testability
- Cost control
- Operational predictability
- Natural fit with human approval
Example flow: contract review
| Step | Step type | What it does |
|---|---|---|
| 1 | Trigger | A document is uploaded |
| 2 | Tool call | The document is parsed |
| 3 | Tool call · knowledge catalogue | Policy documents are searched |
| 4 | LLM call | Risky clauses are extracted |
| 5 | Condition | Is the risk high? |
| 6 · Yes | Human approval | Legal manager approves |
| 6 · No | Tool call · notification | Report and notification |
This table is an example. Document analysis, customer support automation, ticket classification and incident response are built from the same step types.