How to Monitor Harness Runtime Sessionspublic
Last verified 1 Oct 2026
DigitalOcean Harness Runtime provides managed, hardware-isolated microVM sandboxes with built-in tools such as Chromium to run harnesses and execute arbitrary code. Rich lifecycle APIs preserve conversational history and working state across sessions, letting you pause, resume, and fork work to control costs and adapt to the nonlinear nature of agentic workflows. Scale complete agents such as Claude Code or use sandboxes independently for code execution, all through the same service. See What You Can Build for example use cases.
Harness Runtime publishes metrics and lifecycle logs into DigitalOcean Insights, scoped to your team. Use them to see how sessions start, pause, wait for approval, consume tokens, and use sandbox resources, without reading one transcript at a time. Sandbox resource charts reflect the microVM that you can use to right-size a session.
Metrics and logs in Insights are an operational view and not a real-time activity feed that doctl and the Control Panel show while a run is active.
Token metrics displayed are approximate operational signals. Use billing records for exact spend. LangGraph sessions do not publish token data. An empty token chart for that adapter looks the same as an idle session.
View Telemetry for a Session
- In the Control Panel, open Managed Agents and go to Harness Runtime.
- Find the session you want to inspect.
- Click … next to the session and select View insights.
Insights data can lag live activity, so a session you started or paused moments ago may not appear in charts immediately.
Review Available Metrics
Insights covers the following for Harness Runtime:
| Category | What you can see |
|---|---|
| Session concurrency | How many sessions are active now, and how many sandboxes have started over time. |
| Startup and resume | How long agents take to boot, how long a paused session takes to restore, and how long sessions run. Pause, resume, and termination counts show the suspend and resume lifecycle in aggregate. |
| Approvals | How often agents request human approval, how those requests resolve, and how long approvals wait for a person. |
| Token usage | Token input and output per inference call, as a distribution rather than a single average. |
| Sandbox resources | CPU, memory, disk, network, and load inside the microVM, so you can tell whether the sandbox size fits the workload. |
| Coding agent metrics | Selected metrics forwarded from Claude Code and Codex CLI instrumentation. Prompt text and tool arguments from those CLIs are not forwarded. |
| Session origin | Which client created the session (CLI, web, or API) and which region the create request came from. This is recorded at session creation, not who is attached later. |
Lifecycle moments also appear as customer-visible logs in Insights, including sessions created, paused, resumed, and destroyed, and approval requests and resolutions. Approval request text in those logs is truncated.
Use Approval Wait Time to Find Process Bottlenecks
When agents seem slow, check approval wait time before assuming the model or sandbox is the cause. A long wait often means the session is sitting on an unread approval queue, not that inference is slow.
If the policy asks more often than it needs to, pre-approve routine actions. See Configure Agent Permissions.