Agent runs and artifacts
Understand adaptive execution, live run state, tool outputs, reports, and failure recovery.
Adaptive execution
Execution is not forced into three or five rigid levels. The workspace estimates the minimum reasoning budget from the requested consequence, quantitative complexity, available evidence, attachments, and whether the user asked for planning or implementation.
Tool routing is separate from reasoning budget. Only tools allowed by the account tier are exposed to the model, so a Basic response cannot imitate a paid simulation or report export.
Run lifecycle
- 1
Queued
The request and current attachment IDs are durably recorded before background execution starts.
- 2
Running
Events describe meaningful progress. The client can reconnect without inventing completion from a lost connection.
- 3
Completed
The final answer, artifacts, files, and verification state are attached to the run record.
- 4
Failed or cancelled
The run preserves the error boundary. A generated file is not offered as complete when its execution or verification failed.
Artifacts and reports
- Artifacts contain structured analysis results, provenance, limitations, and the action that produced them.
- Generated files are attached only after execution and output checks complete.
- A report is a presentation layer over supported project evidence; producing a polished PDF does not upgrade the underlying confidence.
- Reruns and scenarios preserve lineage so changed assumptions can be compared with the original result.