Observability

An agent is only safe to run in production if you can see what it did and why. The Observe layer gives you complete visibility into every agent decision path — every run, every step, every cost.

#Overview

Observability is the Observe layer of the control plane. Where governance decides what an agent is allowed to do, observability records what it actually did. Every run an agent or pipeline executes is observable: you can trace its reasoning step by step, watch it live as it happens, and review its history long after it completes.

The platform captures the full decision path — the model calls it made, the tools it invoked, the points where it escalated to a human, and the outcome it reached — so you can reconstruct exactly what happened without instrumenting your agents yourself.

#Runs & traces

Every run produces a step-level execution trace. A trace is the ordered record of what the agent did during that run: each model call, each tool call, each escalation to a human, and the final outcome. You can open any run and walk its trace top to bottom to see how it reached its result. See Core Concepts for the run model.

Runs stream live into the control plane. While a run is in progress, its events arrive in real time — you watch the agent reason, call tools, and pause for approval as it happens, not after the fact. A run that is working, a run that is waiting on a human, and a run that has finished are all visually distinct.

• Step-level traces — model calls, tool calls, escalations, and outcome, in execution order.
• Live streaming — in-progress runs surface their events in real time as they happen.
• Run history — completed runs are retained and inspectable, with status, duration, and what triggered them.

#Metrics & cost

Alongside traces, every run emits structured metrics. The platform aggregates them so you can see how your agents behave in aggregate, not just one run at a time: run counts, success and failure rates, and execution duration over a time window.

Token usage is tracked per run and rolled up per agent. Because you bring your own model keys, the platform does not bill you for tokens — instead it gives you an estimated cost for each agent so you can see where spend concentrates and budget accordingly. Cost and usage are broken down by agent and charted over time.

#Incidents

The platform surfaces operational issues across run activity — repeated failures, circuit-breaker trips, and operator-flagged behavior — so a degrading agent is caught early rather than discovered in a postmortem.

When something needs attention, it becomes an operational incident: a tracked record with severity, status, and a timeline, so the response is coordinated and the resolution is documented rather than lost in chat.

#Audit trail

Observability tells you what happened operationally. For the compliance record — the immutable account of every governance decision and agent action — the platform writes a separate, tamper-evident audit trail. Every decision that governs a run is recorded there in addition to being observable here.

The two are complementary: use observability to operate and debug, and the audit trail to prove. See Audit & Compliance for what is recorded and how it is protected.

#Next steps

• Audit & Compliance — The immutable, tamper-evident record of every governance decision and agent action.
• Governance & Policies — How policies decide whether each agent action is allowed, denied, or escalated.
• Deployment — Bring-your-own-keys, environment promotion, and how projects ship to production.