TL;DR
- Langfuse Core is our default for a new, mixed-framework GTM agent: $29/month, shared access and a 90-day history, with tracing, prompts and evaluation in one platform.
- Choose LangSmith when the team already builds in LangGraph or LangChain. Its trace, feedback and evaluation workflow is a close fit; other frameworks can also use it.
- Choose Arize Phoenix when engineers want to own the telemetry and experiment loop. The software is open source; running a production service still costs infrastructure and operator time.
- Choose Helicone when model routing, request cost and latency are the immediate problem. Sessions and custom tool logging extend its view beyond model calls, but connecting those events takes instrumentation.
- Monitor the CRM result as well as the model. A fast, inexpensive run can update the wrong account successfully.
What a GTM trace must explain
A useful production trace answers one question: what happened to this account, and why? Give each research or enrichment job its own run ID, then connect retrieval, model decisions, CRM lookup, proposed fields, the write and its receipt. Keep the company domain and external record ID alongside model and prompt versions.
That turns a disputed qualification into an inspectable chain. The operator can see whether the agent retrieved the wrong business, misunderstood a source, selected the wrong CRM record or lost the response after a successful write. A chat transcript shows only part of that sequence.
This guide compares the monitoring layer. For the execution architecture, see running GTM from a coding agent. The recommendations use documented capabilities and an editorial workload, rather than claiming a comparative product trial.
The four monitoring tools compared
| Tool | Strongest reason to buy | Practical starting price | History basis | Important distinction |
|---|---|---|---|---|
| LangSmith | Integrated tracing, feedback and experiments for an existing agent stack | Developer $0, one seat; Plus $39/seat/month | Base traces 14 days; extended 400 days | One trace can contain many model and tool steps; seats and optional services add cost |
| Langfuse | Framework-flexible tracing, prompt versions and evaluation; cloud or self-hosted | Core $29/month with 100,000 units | Core 90-day data access; Pro $199/month, three years | A unit is a trace, observation or score, rather than one completed job |
| Arize Phoenix | Own an OpenTelemetry-based tracing and evaluation environment | Open-source software; hosting and maintenance separate | Managed by your deployment | Phoenix and the paid Arize AX cloud service are different buying routes |
| Helicone | Model gateway plus cost, latency, sessions and request analytics | Pro $79/month plus usage; Hobby free within limits | Hobby seven days; Pro one month; Team three months | Gateway calls alone do not capture an uninstrumented CRM operation |
Public prices and plan scope were checked on 30 September 2026, before tax. LangSmith pricing, Langfuse pricing, Arize pricing, Helicone pricing.
Capabilities and practical fit
1. LangSmith: best for an existing LangGraph team
LangSmith connects application traces to dashboards, alerts, feedback and online evaluation. Its integrations and SDK support extend beyond LangChain, so choosing it does not require rewriting an application into that framework. The practical advantage for a LangGraph team is that tracing sits close to the graph the developers already debug. Observability documentation.
Use the trace to inspect retrieval and tool branches, attach an incorrect-company label, and turn the failure into an evaluation example. That is more valuable than a dashboard showing aggregate model errors. A sales operator and engineer can discuss the same account run rather than separate screenshots.
Developer includes 5,000 base traces monthly for one user; Plus includes 10,000 and costs $39 per seat. The live pricing table now meters additional base traces at 0.005 LSU each, with an extended-retention upgrade of 0.0025 LSU; one LSU is $1. Those current rates are the basis below. Evaluation compute and other LangSmith services are additional. Pricing and retention.
Non-fit: a small team mainly seeking cheap shared storage across several frameworks may prefer Langfuse Core. Four Plus seats cost $156 before extra trace volume; avoid buying deployment or automated analysis merely to inspect CRM traces.
2. Langfuse: best default for a mixed-framework workflow
Langfuse provides nested application tracing, prompt management, datasets, evaluation scores and dashboards. It supports SDKs and OpenTelemetry, with a self-hosted option. That fits an agent assembled from a coding workflow, custom search and CRM tools without making one agent framework the centre of the monitoring design. Tracing overview.
Core's shared access and 90-day window are useful for monthly sales operations reviews. Pro extends access to three years and adds retention management. The pricing page's broad “unlimited history” wording should be read through its explicit three-year data-access entitlement, rather than as permanent cloud storage. Plan comparison, retention documentation.
The main economic trap is counting jobs instead of telemetry. Every stored trace, observation and score contributes a unit, including entities created by evaluation features. Deep tool instrumentation and several scores per account increase volume even when the number of accounts stays constant. Billable units.
Non-fit: self-hosting to save a $29 subscription rarely makes sense when the team lacks an operator. Buy the managed service unless ownership of the deployment is itself valuable. Native integration with an existing LangGraph workflow can also make LangSmith the lower-effort choice.
3. Arize Phoenix: best for engineers owning the evaluation loop
Phoenix combines tracing, datasets, annotations, experiments and prompt iteration. OpenTelemetry and OpenInference instrumentation feed spans into its collector. The modular SDK offers client, tracing and evaluation packages, including manual tool instrumentation. Phoenix overview, SDK documentation.
This is a strong fit when an engineer wants to compare qualification prompts on known accounts, inspect the failed examples and own the resulting dataset. The same evidence can support a later model change. Instrument the account lookup and write alongside the model; auto-instrumenting an AI library does not automatically explain a custom CRM adapter.
Treat Phoenix as an operated service if it is used in production: database storage, backups, upgrades and trace deletion need an owner. Its software price does not establish an all-in hosting price or a fixed retention period. Arize's current commercial page positions AX separately: AX Pro starts at $50/month with 50,000 spans, 10 GB ingestion and 30-day retention. That is a managed alternative, rather than a Phoenix licence fee. Arize pricing.
Non-fit: a sales operations team wanting a managed dashboard this week should start with Langfuse or LangSmith. Choose Phoenix for control and experimentation capacity, rather than assuming free software removes operating work.
4. Helicone: best for model operations and gateway visibility
Helicone is attractive when the agent already sends calls through a central gateway and the team needs request-level cost, latency, caching or fallback visibility. Its sessions group related calls using session IDs and hierarchical paths. They can include logged vector queries and tool calls, as well as model requests. Session documentation.
That makes it more capable than a model-only usage dashboard. However, a successful gateway request cannot prove a CRM write succeeded: add tool logging and the final receipt to the session. Use one session per account job, rather than a single reused session ID for every customer.
The current paid entry point is Pro at $79/month, with unlimited seats, alerts and reports, plus request/storage usage. Hobby has one seat, 10,000 monthly requests and seven-day retention. Pro's one-month history may cover immediate diagnosis but is short for reviewing an older sales dispute; Team's three-month retention carries a $799 base. Helicone plans.
Non-fit: do not buy Team simply because the gateway is convenient if the primary requirement is inexpensive quarterly account-history review. Langfuse Core provides a longer window at a much lower base. Helicone remains useful where routing and model spend are the central operational problem.
A trace and failure example
Consider a fictional job for northstar-controls.example, account ID A-204. The agent researches recent hiring and proposes a qualification note. Its trace should look like this:
| Step | Evidence recorded | Failure it reveals |
|---|---|---|
| Account job | Run ID, expected domain, CRM ID, prompt version | Jobs or companies mixed under one identifier |
| Search and page retrieval | Query, selected URL, source date and evidence passage reference | Similarly named company selected |
| Two model calls | Inputs by reference, outputs, tokens and duration | Unsupported inference or an expensive reasoning loop |
| CRM lookup | Match rule, returned ID and returned domain | Name-only lookup resolves to A-992 |
| CRM update | Intended ID, field map, attempt ID and response | Correct note sent to the wrong record |
| Receipt | Read-back value and record ID, final outcome | Local timeout hides a completed remote write |
Suppose search, model generation and CRM update all return success, but the returned account domain belongs to another Northstar. An HTTP error alert misses the problem. A deterministic evaluation comparing expected domain and target record identifies it. The operator can then correct the lookup rule and repair the affected record from the saved receipt.
In a second failure, the CRM writes the note but the client times out before seeing the response. Log the attempt and read-back outcome under the same job. The execution layer should reconcile the existing result before repeating the mutation. Monitoring helps explain that recovery; it does not provide the write's idempotency automatically.
LangSmith, Langfuse and Phoenix can represent this instrumented chain. Helicone can group it through sessions and tool logging; two model requests alone leave the lookup and receipt invisible. Store safe field names, IDs and evidence references rather than credentials or unnecessary full contact payloads.
Evaluate business correctness
Start with 50 labelled account cases: 30 straightforward matches, ten similarly named businesses, five old hiring signals and five timeout/retry cases. This is a proposed test set, not a reported vendor benchmark.
Use code checks for the expected CRM ID, allowed field names, source age and exactly one final write outcome. Use a human label or a calibrated model judge for whether the qualification note is useful and supported by its cited passage. A fluent note should still fail if it cites the wrong business.
Our initial release target would be zero wrong-record writes or duplicate outcomes in that set, with at least 45 of 50 notes judged usable. Compare prompt versions on the same examples, then add real production failures to the dataset. Those thresholds are editorial operating targets, rather than a guarantee of accuracy outside the sample.
In production, track accepted notes per completed job, wrong-company incidents, retry rate, unresolved receipts, time to diagnosis and cost per accepted note. Tokens and latency help locate a problem; these outcome measures tell the buyer whether the automation is paying off. Build the account criteria from the signal-based outbound playbook so the evaluation reflects the sales motion.
Cost and retention for 10000 account jobs
Assume 10,000 monthly jobs. Each has one root operation, two model calls and three tool operations: six spans or observations per job. Add one stored business-correctness score per job. There are 20,000 model calls, 60,000 spans and 10,000 scores. Extra retries and evaluation runs would increase those counts.
| Choice | Workload converted into its unit | Monthly monitoring basis |
|---|---|---|
| LangSmith Plus, two users | 10,000 complete traces; 10,000 base traces included | $78 in seats; upgrading 1,000 selected traces at $0.0025 adds $2.50, giving $80.50 before evaluation/other service charges |
| Langfuse Core | 10,000 traces + 60,000 observations + 10,000 scores = 80,000 units | $29, within the 100,000 included units |
| Phoenix, self-hosted | 60,000 spans plus stored evaluation results | Software has no usage fee; illustrative $40 infrastructure + two operator hours at $75 = $190/month |
| Helicone Pro | 20,000 model requests, plus any logged tools and retained payload storage | $79 base plus metered usage; its published calculator depends on request and storage volume, so this is a price basis rather than an invented total |
For LangSmith, one job is one trace only when its steps are properly nested; separately rooted calls, experiments and retries can change the bill. Its current live LSU tariff is used here rather than older $0.50-per-1,000 base-trace references. For Langfuse, doubling this workload to 160,000 units adds 60,000 × $8/100,000 = $4.80 at the first overage tier, giving $33.80. LangSmith price table, Langfuse unit definition, Langfuse rates.
The Phoenix infrastructure figure is a planning assumption, not an Arize quote. Self-hosting may still win when an existing platform team absorbs most maintenance. Likewise, Helicone's two-call count understates full-chain logging if additional tool events are submitted. Model generation, search, CRM subscriptions and evaluation-model tokens remain separate from the monitoring budget.
Retention changes the decision. LangSmith can keep most traces for 14 days and extend selected failures for 400. Langfuse Core makes the full 90-day access window easier for recurring operations review. Phoenix lets the operator set policy and capacity. Helicone Pro's one month favours rapid model diagnosis over long account-history investigations.
The meaningful return is investigation time. If instrumentation reduces 20 monthly incidents from 25 to ten minutes each, it saves 300 minutes, or five hours. At an assumed $75/hour, that is $375. Against Langfuse's $29 base, the remainder is $346 before implementation and other costs. An eight-hour initial setup at the same rate costs $600; that example recovers setup in about 1.7 months if the savings persist. These are explicit assumptions, rather than measured product results.
Our recommendation
Start a new mixed-framework account-research agent on Langfuse Core. Its shared access, useful history and integrated evaluation cover the practical need at a low base. Instrument the lookup and write receipt before expanding telemetry detail.
Stay with LangSmith when the developers already use its traces and evaluation workflow around LangGraph. The avoided integration and context-switching work can be worth more than a smaller storage bill. Choose Phoenix when an engineering team wants to operate the platform and own the experimentation environment. Choose Helicone Pro when gateway routing and model cost diagnosis dominate, while explicitly logging the tools needed to explain the business result.
The deciding feature is whether an operator can reconstruct one disputed account update. A dashboard that reports inexpensive successful calls without the correct record identity is insufficient for that job.
FAQ
Is LangSmith limited to LangChain?
No. It supports other frameworks, providers and SDK instrumentation. An existing LangGraph team is a particularly close practical fit.
Is one account job one billable unit?
Not across vendors. LangSmith can nest several steps in one trace; Langfuse counts traces, observations and scores. Helicone counts logged requests and storage. Instrumentation determines the actual volume.
Does Phoenix have the same pricing as Arize AX?
No. Phoenix is the open-source platform. AX is the separately priced managed commercial service. A self-hosted Phoenix budget includes infrastructure and operation.
Can Helicone show tool calls?
Yes, through logging grouped into sessions. Passing only model calls through the gateway will not automatically expose separate CRM operations.
Which metric matters most for a sales-research agent?
Cost per accepted, source-supported account result. Track wrong-record writes and unresolved receipts alongside it so apparently efficient runs do not hide incorrect outcomes.
