Execution Monitoring and Logging
Execution list, live event stream, logs / cost / tokens, and error localization—see the run state clearly.
Once a workflow is running, whether you can make sense of "what it actually did and how much it cost" determines whether you'll dare put it into production. This page covers the execution list, execution details, real-time events, token and cost records, export, and the data dashboard.
Execution List Page
Left nav → "Execution history." It supports card and table views, and the list auto-refreshes (pausable) while executions are running. Filtering and sorting:
- Keyword search — by execution ID or workflow name
- Status filter — PENDING / RUNNING / COMPLETED / FAILED / CANCELLED / INTERRUPTED
- Workflow filtering - only view the history of a certain workflow
- Sort — start time / finish time / duration / workflow / status / progress
At the top of the list is a row of quick filters with counts (Running / Failed / Completed / Interrupted); click to toggle.
Information Displayed for Each Record
- Status badge + execution ID + workflow name
- A progress bar and the step completion count (e.g. 5/7)
- Start time (relative) and duration
- RUNNING records also show the step currently running
- Finished records can be exported to JSON / YAML inline
After multi-selecting, you can bulk-cancel running executions or bulk-delete finished records.
PENDING— Created but not yet started. Executions over the concurrency limit queue in this state and automatically move to RUNNING once earlier ones finish.RUNNING— At least one step is executing.COMPLETED— All steps have finished (including steps SKIPPED because their condition wasn't met).FAILED— A step errored, and the execution ended in failure.CANCELLED— Manually cancelled by the user.INTERRUPTED— Interrupted by a system cause such as a service restart. For whether it auto-resumes, see the auto-resume page under "Next steps" below.
Execution Details Page
Click any execution to open its details page. From top to bottom:
- Top bar — cancel, refresh, export (JSON / YAML), and jump to the corresponding workflow editor
- Status card — status, progress, duration, and a token overview (sent / received / total)
- Main-area tabs (see the table below)
- The bottom "Execution log" stream — a scrolling log of all events, filterable by level (debug / info / warn / error), searchable by keyword, with auto-scroll pausable
| Tab | Content |
|---|---|
| Steps | A vertical timeline: each step's status, duration, and iteration progress. Click to open the step details side panel; after execution ends, you can rerun from a chosen step. |
| Chat | Weaves Agent utterances, LLM output, tool calls, and system messages into a single chronological chat stream. |
| External Messages | See only the IM channel send/receive records (Slack / Telegram, etc.). |
| DAG | A step graph drawn by dependency, colored by run status in real time. |
| Playback | Replay this execution along a timeline. |
| Chain Of Reasoning | The Agent's intermediate reasoning. |
| Artifact | All artifact files: preview, download, share a single file, or generate a public link for the whole directory. |
| Performance | Post-execution duration analysis. |
| Debug | Appears when breakpoints have been set or a debug session exists. |
Step Details Side Panel
Click a step in the timeline and a side panel slides out with everything about it: Output, Events, Reasoning, Artifacts, and Conversation tabs (iteration / interaction / orchestration steps get extra tabs), plus the start and end times and a "Rerun from this step" button. A failed step's error message and retry count are here too.
Real-Time Updates (WebSocket)
While an execution is RUNNING, the details page receives incremental updates over WebSocket without manual refresh. A dropped connection reconnects automatically (exponential backoff, up to 5 times); if that still fails, it falls back to polling every 2 seconds and shows a notice.
Push Message Types
| Type | Meaning |
|---|---|
full_status | A full snapshot of the execution status |
status_change | Execution status / progress changes |
step_update | A status update for a single step |
new_event | A new execution event (see event types below) |
artifact_found | A new artifact file was generated |
user_interaction_request | A step requests human input, and a reply box appears on the conversation tab |
List Of Event Types
Each step's event stream uses the type field to distinguish event kinds. The common ones, listed by family:
| Event Type | Meaning |
|---|---|
agent_starting / agent_completed / agent_failed | The lifecycle of an Agent step |
llm_call_starting / llm_call_completed | One LLM call; the completed event carries this call's token count |
llm_streaming_starting / llm_streaming_completed / llm_streaming_failed | Streaming LLM output |
tool_call_starting / tool_call_completed / tool_call_failed | An Agent tool call, with arguments and a result summary |
external_message / im_message / telegram_message | IM channel message send/receive |
group_chat_started / group_chat_message / group_chat_summarizing / group_chat_completed | The multi-Agent discussion in a group_chat step |
state_machine_started / state_machine_transition / state_machine_state_completed | The state transitions of a state_machine step |
agent_based_started / agent_based_completed | The delegation process of an agent_based orchestration step |
user_interaction_request / user_interaction_message / user_interaction_timeout | Asking the user a question mid-execution, receiving a reply, and waiting for a timeout |
claude_code_* | Streaming progress events from the Claude Code runtime |
Events also carry a category field (agent / llm / tool / message / strategy / user_interaction), and the conversation tab's category filter groups by it.
Conversation-View Filters
- Keyword search — searches message content and the speaker
- Filter by Agent — see only a given Agent's or step's messages
- Category filter — Agent / LLM / Tool / External message / Strategy / Interaction
- Auto-scroll toggle — whether to follow to the bottom as new messages arrive
Tokens And Cost
- LLM-related events record input / output tokens per call, and the status card sums them live to show sent / received / total.
- After execution ends,
GET /api/executions/{executionId}/costestimates this execution's cost using the prices of the models used, returning the total and a per-step breakdown. - Cost is estimated afterward from token usage; the platform doesn't quote costs before a run.
Export Execution Process
The "Export" action in the details top bar or a list row exports an execution to a JSON or YAML file containing the execution status, results, performance metrics, and a snapshot of the workflow definition and YAML source at export time. Sensitive fields like passwords, API Keys, and webhook addresses are automatically replaced with [REDACTED]. The file name looks like execution-process-79621f70-completed.json.
The workflow definition in the export is the current version pulled at export time; if the workflow was changed after this execution, it may differ from the definition at the time of the run.
Data Dashboard
The cross-execution aggregate view is in Analytics the dashboard: total executions, success rate, average duration, and total token usage, plus execution trends, status distribution, and a top-workflows ranking, with a time range of 24 hours / 7 days / 30 days.
Error Location
When an execution is FAILED, first find the red step in the timeline and open its details: the error message is on the "Output" tab, and the process details are on the "Events" tab (a code step's stderr is here too). Steps that were retried show their retry count.
For how to handle common errors, see Troubleshooting.
Retention Period
Execution records (status, results, events, performance data) and artifact files are retained for 90 days by default, after which the platform cleans them up on a schedule; the exact period depends on your plan or deployment config, and a running execution is never cleaned up.
Next
- Breakpoint Debugging — Use it when you "can't understand why you are running like this"
- Troubleshooting — Common errors and how to fix them
- Auto-Resume — How to handle INTERRUPTED executions