Agent and sandbox telemetry
Record agent runs, sandbox operations, commands, and application output as queryable telemetry.
Agent and sandbox telemetry
Use RawTree as a flight recorder for agent and sandbox runs. Stream sandbox operations, commands, file changes, process output, and application telemetry into RawTree, then reconstruct what happened with SQL or an agent.
This is useful when an autonomous run fails, succeeds for the wrong reason, or needs to be audited after the sandbox is gone. Instead of relying only on terminal scrollback or provider logs, keep each operation as queryable event history.
Where it fits
Agent and sandbox telemetry helps when you need to understand behavior across many runs:
- Which command failed, and what happened immediately before it?
- How long do sandbox setup, dependency install, tests, or teardown steps take?
- Which files changed during a run, and which tool or command touched them?
- Did the application inside the sandbox emit traces, logs, or errors at the same time?
Send control-plane events from the sandbox provider, command-level events from the agent harness, and application telemetry from workloads running inside the sandbox. RawTree can store each stream directly and query them together by run ID, sandbox ID, trace ID, or timestamp.
What to store
Start with enough context to replay the run:
- Run, task, sandbox, workspace, and user identifiers.
- Command start, exit, duration, stdout/stderr summaries, and exit codes.
- File change events, tool calls, lifecycle events, and provider operations.
- OpenTelemetry traces or logs from applications running inside the sandbox.
Keep payloads structured when possible. Agents can use the same SQL queries that humans use, and structured rows make it easier to summarize failures or compare runs.
Start with
- Choose a stable run ID and include it on every event.
- Send sandbox SDK spans or command events into RawTree.
- Query recent failures by run ID, command, duration, or exit code.
- Add application OTLP telemetry when you need to correlate agent actions with runtime behavior.