What you can check today

Runs execute on a durable workflow engine. Every save creates an immutable version and each run is pinned to the version it started on. Runs can be paused, resumed, cancelled or terminated, a failed node can be retried with its original inputs, and a run can be started again from a chosen node.

Every run shows live per-node state and a streaming run log, including what your own Code nodes print, and each node keeps its inputs and outputs for inspection.

Not there yet: A step-by-step trace of an AI agent’s reasoning and tool calls, and OpenTelemetry GenAI spans you can export to your own tracing tool, are in development.

Live on production: every model call a workflow run makes is recorded with its provider, model, token counts and a server-side price, and can be read per run and per workspace for a date range.

Not there yet: There is no public page yet where you can see this for yourself: the documentation and a published demo run are still to come, and per-call cost is not yet shown alongside an agent’s step-by-step trace.

Not available yet. FluidGrids does not offer evaluation datasets, scored evaluation runs or evaluation gates on publishing today.

Not there yet: Evaluation datasets, scored runs, an evaluation gate before promotion, and guardrail checks on model input and output.

Workspace action policies allow, deny or require approval for outbound actions, with default deny wherever a policy matches. An AI agent’s tool call can be set to wait for a person’s approval, and promotion to the Production environment needs a second approver by default.

A React SDK embeds the workflow builder in your own product at three integration depths, without an iframe and under the FluidGrids name, and a public read-only viewer renders a workflow definition on any page.