Operate / 10

Inspector and observability

Foundry is the runtime and the workbench. The same correlation model that runs the system explains it: instance → conversation → run → pass → tool, message, transmission, schedule, artifact, and handoff.

What the inspector shows

ViewQuestion it answers
Run summaryWhat was requested, what state is it in, and what outcome or failure was returned?
Agent activityWhich agents are active, waiting, sleeping, retrying, or complete?
Message passingWho handed what to whom, in which conversation, and with which artifact?
Model passesWhich provider pass ran, for how long, and what usage or retry occurred?
Tool and workbench activityWhat capability ran, what safe input/output summary was recorded, and which files changed?
Inbound and outboundWhich transmission matched, which subscription provisioned targets, and where was a reply delivered?

Work notes, not hidden reasoning

The runtime can display declared intent, current action, progress, decisions, outputs, and safe summaries emitted by the agent or subscriber. It does not expose private chain-of-thought. Ask agents to emit useful working notes—“reviewing three directions against the brief”—and keep secrets, raw credentials, and unfiltered provider payloads out of telemetry.

One event envelope

Events carry timestamps plus run, definition, instance, conversation, workspace, pass, and parent correlation where relevant. Custom subscribers can stream them to OpenTelemetry, a warehouse, a product-specific interface, or an audit store without changing the agent definition.

Provider pressure is a state

A transient 429 or interrupted stream should appear as a paused pass with its retry attempt and backoff, not as a completed agent or an unexplained wall of logs. Terminal exhaustion fails the run with the provider status. The inspector distinguishes a retrying pass, a sleeping run, a waiting handoff, and a finished agent.

Embed the event stream

Product interfaces can subscribe to the same event API and render a simpler experience: campaign lanes, agent status, handoffs, artifacts, approvals, and a lead conversation. The Foundry workbench remains the deep diagnostic view; your application decides which operational concepts its users should see.