Locus product guide · Updated November 10, 2026

AI agent observability: tracing, costs and tools

AI agent observability means recording every step an agent takes as a trace: model calls, tool calls, inputs, outputs, latency, tokens and cost. Langfuse Cloud is free for 50,000 units a month, LangSmith Developer is free for 5,000 traces, and Datadog Agent Observability is free up to 40,000 LLM spans. These tools track model cost well. The bills from paid APIs your agent calls need their own call log. Prices checked November 10, 2026.

What should AI agent observability capture?

One agent request can fan out into many model calls, tool calls and retries. A trace ties them together so you can see why the agent did what it did and what it cost.

A trace per run, with a session or conversation ID and a user ID.
A span per model call: model, prompt, output, tokens, latency and cost.
A span per tool call: tool name, arguments, result size, status, latency and any charge.
Errors, retries and refusals, with the reason.
Evaluation scores, from rules, models or human review.
The money spent outside the model: paid APIs, data providers and scrapers.

How do OpenTelemetry and MCP fit in?

OpenTelemetry has semantic conventions for generative AI, now kept in their own repository. They define spans, metrics and events for GenAI clients, for MCP, and for specific providers. Using them keeps your traces portable between backends.

MCP carries trace context too. SEP-414, now final, documents W3C traceparent, tracestate and baggage keys in a request’s _meta field, so a tool call can join the caller’s trace. The 2026-07-28 spec also deprecated MCP’s own logging feature and points servers to stderr or OpenTelemetry instead.

How do agent observability tools compare on price?

Each vendor meters something different, so the same agent costs very different amounts. Figures come from each pricing page on November 10, 2026. Prices checked November 10, 2026.

How do agent observability tools compare on price?
Tool	Free tier	Paid entry	What is billed
Langfuse Cloud	Hobby: 50,000 units a month, 30 days of data, 2 users	Core: $29 a month with 100,000 units, then $8 per 100,000	Units: every trace, span, event, generation and score
LangSmith	Developer: 1 seat, 5,000 base traces a month	Plus: $39 per seat a month with 10,000 base traces, then pay as you go	Seats plus traces. Base traces keep 14 days, extended traces 180 days.
Datadog Agent Observability	Up to 40,000 LLM spans a month, 15-day retention	Pro: $160 a month on annual billing for 100,000 LLM spans, then $3.50 per 10,000	LLM spans only, not tool or workflow spans

What does it cost to trace 10,000 agent runs a month?

Assume each run makes 8 model calls and 12 tool calls, and gets 1 evaluation score. That is 1 trace, 20 spans and 1 score per run.

Langfuse Core: 22 units a run is 220,000 units. $29 covers 100,000, and the other 120,000 cost $9.60, for $38.60 a month.
LangSmith Plus with 1 seat: 10,000 traces fit in the 10,000 included, for $39 a month with 14-day retention.
Datadog: 8 LLM spans a run is 80,000 LLM spans, above the 40,000 free tier. Pro covers 100,000 for $160 a month on annual billing, or $240 on demand.

How do I see what an agent spends on paid tools?

Tracing tools price model tokens well. They do not know what a search, scrape or enrichment call cost, because that bill comes from each provider. Datadog, for example, bills only LLM spans, not tool spans.

Locus Pro records every paid API call your agent makes through it, in one call history. Each entry has the tool, provider, status, credits charged, time and any error. Each HTTP response also returns the credits charged, the balance after the call, and a call ID.

To join that spend to your traces, send your trace ID with the call. On the HTTP route, the X-Locus-Attribution header takes JSON such as a task ID and feature name. Enterprise workspaces can filter the ledger by those keys, or by end customer.

Example charges: Brave web search costs 30.3 credits, Parallel search 1 to 15.2 credits and Hunter email finder 1 to 10.1 credits, where 1,000 credits equal $1. Prices checked November 10, 2026.

Which AI agent observability tool should I choose?

Start from where your team already works and how much data you send.

You want open source or to keep data in house: Langfuse, self-hosted or on its cloud.
You build on LangChain or LangGraph: LangSmith.
You already run Datadog for the rest of your stack: Datadog Agent Observability.
Your agents call many paid APIs and you need the spend per run or per customer: add Locus Pro call history next to any of these.
Locus Pro is not a tracing tool. It does not record model calls or tools you call outside Locus.

Frequently asked questions

What is the difference between agent observability and monitoring?

Monitoring tells you that something is wrong, such as a spike in errors or cost. Observability lets you trace one run step by step to see why, including the prompts, tool calls and results.

What is an LLM span?

Datadog defines it as one call to a model provider such as OpenAI or Anthropic. One agent run can make several. Tool, retrieval and workflow spans around those calls are not billed as LLM spans.

Can I self-host agent observability?

Yes. Langfuse is open source and free to self-host, with Docker Compose for a local setup or Kubernetes templates for production. You pay for the servers and storage.

Does MCP support tracing across tool calls?

Yes. MCP documents W3C trace context keys in each request’s metadata, so a server can continue the caller’s trace. Both the client and the server have to implement it.

Implementation references

Use these first-party references for current request contracts and account requirements. Tool availability, prices, and negotiated terms can change.

Sources

Third-party details were checked on November 10, 2026. Vendors change pricing and features often, so confirm on their site before you decide.