Autonomy you can inspect
An autonomous pipeline is only trustable if you can see inside it when you need to. Forge instruments the runtime so that every turn leaves a readable trail — not just “the run failed,” but why this turn made the call it did.
Logs and metrics, per turn
The worker emits rich instrumentation as it runs — iteration boundaries, each tool call and its result, model responses, errors — into a structured log stream. In the dashboard, each turn gets:
- a Logs tab with level filtering and free-text grep, autoscrolling as the turn runs;
- a Metrics tab with RSS / CPU / iteration-rate sparklines, sampled every few seconds and bucket-aggregated server-side;
- a per-turn Debug toggle that expands the full instrumentation trace when you want the deep view.
One feed for everything
Above the per-turn detail, a unified activity feed gives you a single typed-union surface across workflows, pipelines, chats, and cloud sessions — scoped to your org, filterable by type, and deep-linked so one click takes you from “something’s happening” to the exact run. It’s the operator’s home base for a fleet of agents working in parallel.
The boring reliability work — knowing what every turn did and why — is what makes looking away from the loop safe.