Testing, logs and observability

Every message creates a run, and every run is fully recorded.

Runs. Status, timing, input and output, token usage (actual when the provider reports it, estimated otherwise), cost, the workflow version, and the checkpoint reference for resuming.

Steps. One per node, with its context report: what the node did, how much history and knowledge it kept, what it trimmed and why, and how long it took. Runtime housekeeping steps are shown as internal. A step names a provider and model only when it called one: an LLM step shows the model that actually ran, including a model set on the node itself, and every other step (an app's Chat (in), Agent and Chat (out) steps included) names none.

Edges. What moved from one node to the next.

Generations. Each model call: provider and model, the exact messages sent (system prompt, history, memory and knowledge context, tool results), the raw request settings, the response, finish reason, latency, time to first token, tokens and cost.

Tool calls. Each call with its input, output, status, duration, retries and errors, plus what the model was shown of the result. An event tool's call records that it ran in the chat window, the event, the mode and how long it waited.

Pauses. A step the run paused on — a tool approval, or a question the agent asked — shows what it waited on and how it ended: the questions with their options and the answer (or the cancel, or the timeout), the tool calls and the decision. The run continues on that same step when it resumes, so a pause never splits a node into two steps. The follow-up run of an async event is marked as such and links back to the run that fired the event.

Retrievals. The query, candidates, scores, the selected chunks and their text, what was trimmed for the prompt, and whether a source chunk has changed since.

Memory reads and writes. Which items were considered and selected, and which candidates were extracted.

Conversation storage. The run detail shows how much the conversation occupies as stored, split into logs, checkpoints and shared retrieved text.

Runs can be filtered by app, agent, status and time under Runs. Analytics shows the organisation's last 7, 30 or 90 days against the period before: runs (completed and failed), cost, tokens, conversations and hand-offs per day, average and p95 latency, and breakdowns by agent, app, model, channel and tool. Settings → Plan & usage shows this month's runs and conversations against the organisation's plan (Plans, pricing and what's next).