Orchestrate across systems
The agent does not follow a fixed script. It breaks the task down, sequences steps by dependency, and adjusts when something does not go as expected.
Two execution modes
Standard mode (ReAct) thinks and acts one step at a time, deciding the next step from the last result. It suits work that needs judgement and retries: alert triage, document analysis, information retrieval.
Planner mode (DAG) builds the full plan first, splitting the task into steps and identifying dependencies. Independent steps run at the same time, dependent ones run in order. The result is checked afterwards, and if the goal was not met the plan is revised and retried, up to three rounds.
Automatic routing
By default a lightweight classifier judges query complexity: simple work goes to standard mode, complex multi-step work goes to planner mode.
The interface exposes Auto, Standard, and Planner, so you can pin the behaviour when you need it fixed.
Watching it run
With cross-system work the problem is rarely the final answer. It is one step in the middle, and these are visible while the run is in progress:
- An interactive graph showing node status, dependencies, and how long each step took.
- A full think-act loop inside every step, with one automatic retry on failure.
- Tool-call caching within a run, so the same call does not hit the external system twice.