Built In Metrics

Agent Metrics Instrumentation

You’re using a skill that wires LaunchDarkly agent metrics around an existing provider call. Your job is to audit what’s already there, pick the right tier from the ladder below, and implement it with the least ceremony that still captures the metrics the Monitoring tab needs (duration, input/output tokens, success/error, plus TTFT when streaming).

The single most important thing to get right: default to the highest tier that fits the shape of the call. Going lower (“just write the manual tracker calls”) looks flexible but costs you drift, missed metrics, and legacy patterns the SDKs have moved past.

The four-tier ladder

This is the order the official SDK READMEs (Python core, Node core, and every provider package) recommend. Walk from the top and stop at the first tier that fits:

Every provider — OpenAI, LangChain, Vercel, Bedrock, Anthropic, Gemini, custom HTTP — uses the same generic shape: tracker.trackMetricsOf(getAIMetricsFromResponse, () => providerCall()) in Node, tracker.track_metrics_of(get_ai_metrics_from_response, provider_call) in Python. The extractor is the only thing that changes per provider: import getAIMetricsFromResponse from the matching @launchdarkly/server-sdk-ai-<provider> (or ldai_<provider>) package, or write a small custom function that returns LDAIMetrics. There are no provider-specific tracker methods.

Workflow

1. Explore the existing call site

Before picking a tier, find the provider call and answer these questions:

• Not done: Shape? Is it a chat loop (history + turn-based), a one-shot completion, an agent step, or something else? → drives Tier 1 vs 2.
• Not done: Framework? Raw provider SDK? LangChain / LangGraph? Vercel AI SDK? CrewAI? Strands? → drives which Tier-2 provider package (if any) applies.
• Not done: Provider? OpenAI, Anthropic, Bedrock, Gemini, Azure, custom HTTP? → cross-reference with the package availability matrix below.
• Not done: Streaming? If yes, you’ll need TTFT tracking, which means Tier 4 for the TTFT part even if the rest is Tier 2.
• Not done: Language? Python or Node? Provider-package coverage differs between them.
• Not done: Already using a config? If not, route to configs-create first — tracking requires a tracker, which is obtained by calling create_tracker() / createTracker() on the config object returned by completion_config() / completionConfig() / createModel().
• Not done: On the current SDK API? If the call site uses aiclient.config(...) / aiClient.config(...) or constructs an AIConfig(...) / LDAIConfig default, it’s on the pre-0.20 surface. Migrate it as part of this work before adding tracking:
  • aiclient.config(...) → aiclient.completion_config(...) for one-shot/chat or aiclient.agent_config(...) for agent mode (mirror the call signature). Node is the same with camelCase.
  • AIConfig(...) default → AICompletionConfigDefault(...) or AIAgentConfigDefault(...) (Node: / ). is the base class the SDK returns; it isn’t a valid default-value constructor — the typed variants are.

2. Look up your Tier-2 option

Use this matrix to decide whether Tier 2 (provider package) is available for your situation. If it’s not, drop to Tier 3 (custom extractor). If the shape is chat-loop, go to Tier 1 first regardless of what’s in this matrix.

3. Implement from the matching reference

Once you know the tier and the provider, open the reference file and follow the pattern. The references are written so Tier 1 is always the first example, Tier 2/3 next, and Tier 4 last. Stop at the first tier that matches the app’s shape.

Guardrails that apply to every tier:

1. Always check config.enabled before making the tracked call. A disabled config means the user has flagged the feature off — you should short-circuit to whatever fallback the app uses (cached response, error, degraded path) rather than making the provider call at all.
2. Wrap the existing call, don’t rewrite it. Tier 2 and Tier 3 are designed to slot around an unmodified provider call. If you find yourself rewriting the call to fit the tracker, you’re at the wrong tier — drop down one.
3. Errors are handled inside trackMetricsOf. The wrapper catches exceptions, records trackError() internally, and re-raises — do not add except: tracker.trackError() on top, it’s a noop that also trips the at-most-once guard. Tier 1 handles both paths automatically. At Tier 4 (manual, streaming, track_duration_of) the caller does own the error-tracking call.
4. Always flush before close. Call ldClient.flush() (Python: ldclient.get().flush(); Node: await ldClient.flush()) before closing the client. Trailing events are at risk of being lost otherwise — in short-lived scripts and long-running services alike. In Node, ldClient.close() returns a Promise; await it.

4. Verify

Confirm the Monitoring tab fills in:

• Not done: Run one real request through the instrumented path.
• Not done: Open the config in LaunchDarkly → Monitoring tab. Duration, token counts, and generation counts should appear within 1–2 minutes.
• Not done: Force an error (bad API key, zero max_tokens, whatever) and confirm the error count increments.
• Not done: If streaming: verify TTFT appears. If it doesn’t, you probably wrapped the stream creation with trackMetricsOf but didn’t add the manual trackTimeToFirstToken call — see streaming-tracking.md (opens in a new tab).

Quick reference: tracker methods

Obtain a tracker via the factory on the config object: tracker = config.create_tracker() (Python) or const tracker = aiConfig.createTracker() (Node). Call the factory once per execution and reuse the returned tracker for every call — each factory invocation mints a new runId that tags every tracking event emitted by that tracker so events from a single execution can be correlated together (via exported events / downstream systems). The Monitoring tab aggregates events rather than grouping them by run today — the runId is useful when events are exported or queried outside the UI, and is the identifier the SDK’s at-most-once guards are keyed on. The methods below are the raw API surface — most of the time you should not call them individually; use trackMetricsOf or a Tier-1 managed runner. The list is here so you can recognize the methods in existing code and reach for the right one when you genuinely need Tier 4.

Related skills

• configs-create — prerequisite if the app doesn’t have a config yet
• custom-metrics — business metrics (conversion, resolution, retention) layered on top of the agent metrics this skill captures
• online-evals — automatic quality scoring (LLM-as-judge) on sampled live requests; complementary to the metrics here
• migrate — Stage 4 of the hardcoded-to-AgentControl migration delegates to this skill