If you can't trace it, you can't trust it
Every run leaves a complete, secret-free trace — and runs with a price tag attached. Observability and cost control are runtime features, not add-ons.
Traces
Full visibility into every run
Every run is a trace
ShippedModel calls, tool calls, inputs, outputs, latencies — each run emits standard OpenTelemetry spans linked from the run record. Nothing about agent behavior is a black box.
Langfuse inspection
ShippedTraces land in Langfuse for step-by-step inspection, scoring and debugging. Operators open the trace of any run in one click from the console.
No tokens in the logs
ShippedCredentials and JWTs are redacted at instrumentation level before traces are stored — observability never becomes an exfiltration path.
Self-hosted trace store
GA Q3 2026Langfuse runs as a managed cloud or self-hosted inside your boundary — traces can stay wherever your data residency policy requires.
What a trace holds
Anatomy of a governed run
Everything needed to answer 'why did the agent say that?' — months later, to an auditor.
trace · aro.reconciliation-investigator · run #4821
- descriptoraro.reconciliation-investigator@1.4.2 · catalog sha 9f31c2e
- modelclaude (via gateway) · 14,302 tokens · $0.21
- tool_call 01aro.get-break-context — 412 ms · ok
- tool_call 02aro.get-match-history — 388 ms · ok
- tool_call 03dh.get-security-master — 201 ms · ok
- outputschema-valid · root_cause + evidence[3] + recommended_action
- auditappend-only entry · approver n/a (task mode) · retained 7y
Cost
Bounded by design, attributed by default
Hard ceilings per run
ShippedMax tokens, max tool calls, max duration — declared in the descriptor and enforced by the runtime. A misbehaving run stops itself.
Cost attributed per run
ShippedToken usage and model cost are recorded on every run record — cost is a first-class field, not a monthly surprise.
Tenant budgets & rate limits
BetaPer-tenant rate limiting is live; workspace-level budget envelopes and alerts arrive with the multi-tenant GA.
Eval scores over time
BetaRun outcomes feed evaluation baselines, so quality is observable the same way cost is — per agent, per version, over time.
FAQ
Frequently asked questions
Can our own observability stack consume the traces?
Yes. Runs emit standard OpenTelemetry spans, so they can be exported to any OTel-compatible backend. Langfuse is the supported inspection UI, but the trace format is open.
What exactly is recorded for a run?
The descriptor version (catalog SHA), model identifier, ordered tool calls with inputs and outputs, token counts, cost, duration, final status, and the structured output. Together with the audit log this reconstructs any run end to end.
Do traces contain sensitive data?
Traces contain the operational payloads agents worked with, under the same tenant isolation as the rest of the platform — but never credentials: tokens and secrets are redacted before storage. Self-hosting the trace store at GA keeps traces inside your boundary entirely.
How do you prevent runaway LLM costs?
Three layers: per-run ceilings (tokens, tool calls, duration) enforced by the runtime, per-tenant rate limits, and cost recorded per run for attribution. Budget envelopes with alerts ship with multi-tenant GA.
Is observability optional?
No — tracing is built into the runtime rather than offered as an integration. A run without a trace would be ungoverned by definition; that path does not exist.
Keep exploring
More of the platform
- Platform overviewThe governed agentic platform, end to end.
- Agents & catalogDeclared, versioned agents — descriptors, skills, modes.
- Governance & evalsPR review, eval baselines, allow/ask/deny policies.
- Security & trustIdentity, tenant isolation, vaults, append-only audit.
- DeploymentEU SaaS today; hybrid, on-prem and sovereign at GA.
- Integrations & MCPMCP tools, NeoXam products, open standards, models.
- ArchitectureRunner, harness, control plane vs execution plane.
General availability comes in Q3 2026. The Early Adopter Program is open now.
A limited cohort, a one-year platform trial and three workshop streams — Business ROI, Compliance, Operational fit. Bring one workflow; leave with a governed agent and the evidence to certify it.