Langfuse+46 Positive

Operator Insights

What are teams learning from building on Langfuse and running it in production?

17 accounts: 2 left it or turned it down, 4 switched, 1 evaluated it and 10 adopted it

Left it or turned it down2

Jérémie DoucyContributor, bayes-platformBayes Impact
Bayes Impact's bayes-platform, in staging and production, exporting LLM traces over OpenTelemetry.
GitHub (bayesimpact/bayes-platform), published Sep 29, 2026Read the write-up (opens GitHub (bayesimpact/bayes-platform) in a new tab)
Observation
Bayes Impact is leaving Langfuse: a merged pull request has its platform emit plain OpenTelemetry, and a collector sends LLM traces to Arize Phoenix.
Takeaway
Export plain OpenTelemetry and let a collector pick the backend, so leaving Langfuse is a collector change, and plan the replacement's production database and authentication before you move.
What happened
A pull request merged on September 29, 2026 deleted the platform's 673-line Langfuse exporter, the langfuse packages and two Google Cloud OpenTelemetry exporters. The platform now emits standard OTLP and knows no trace backend; an OpenTelemetry Collector maps spans to OpenInference and sends them to Phoenix. The plan keeps Langfuse up, read-only, for 30 days, until both hold the same trace history, and the Cloud Run production environment has no gateway yet and exports nothing with this version. The same pull request notes that the Phoenix bundled in its Helm chart is documented as not production ready, with no authentication and SQLite by default.
“…the platform emits standard OTLP and knows no trace backend. The OpenTelemetry Collector does the mapping to OpenInference and sends the LLM traces to Phoenix.”
Jérémie Doucy
- Public engineering write-up (GitHub pull request)
- Code change merged
Deployment details
Bayes Impact's bayes-platform, in staging and production, exporting LLM traces over OpenTelemetry.
- Public engineering write-up (GitHub pull request)
- Code change merged
Team
Bayes Impact platform engineering
Workload
LLM traces from the bayes-platform services
Product version
Langfuse exporter and langfuse packages, removed on September 29, 2026
In SignalsCounted once, as the contributor's own pull request. In the record

digithings.aiOct 1, 2026GitHub (digithings-ai/digithings)digithings.ai approved self-hosted Langfuse to replace LangSmith, then chose Cloudflare-native tracing with MLflow instead.
Chris (chrizefan)Owner and approver of tooling decisionsdigithings.ai
LangGraph workloads behind a pluggable OpenTelemetry layer, with LangSmith in place.
GitHub (digithings-ai/digithings), published Oct 1, 2026Read the write-up (opens GitHub (digithings-ai/digithings) in a new tab)
Observation
digithings.ai approved self-hosted Langfuse to replace LangSmith, then chose Cloudflare-native tracing with MLflow instead.
Takeaway
Put an OpenTelemetry layer in front of the tracer first; it let one team approve Langfuse and reverse the choice eight days later without touching its agents.
What happened
On October 1, 2026 the owner approved self-hosting Langfuse on Cloudflare to replace LangSmith, with the same spans sent to both until a parity run. On October 9, 2026 the plan changed: digithings.ai chose Cloudflare-native tracing with MLflow instead, to avoid adding paid services, and kept LangSmith until the parity run.
Public engineering write-up (GitHub issue)
Decision record
Deployment details
LangGraph workloads behind a pluggable OpenTelemetry layer, with LangSmith in place.
Public engineering write-up (GitHub issue)
Decision record
Team
digithings.ai engineering
Duration
Approved October 1, 2026; reversed October 9, 2026
Workload
LangGraph agent workloads
Product version
Self-hosted Langfuse on Cloudflare, planned and dropped
In SignalsCounted once, as the owner's own issue. In the record

Switched to or from it4

mfundnContributor, bbvch-ai/aihub-corebbv
bbv's AI Hub core, with agent tracing, datasets and automatic provisioning of the trace backend.
GitHub (bbvch-ai/aihub-core), published Feb 12, 2026Read the write-up (opens GitHub (bbvch-ai/aihub-core) in a new tab)
Observation
bbv's AI Hub replaced Arize Phoenix with Langfuse in February 2026, moving its tracing, datasets and trace-backend provisioning.
Takeaway
Scope a move from Phoenix to Langfuse across tracing, datasets and provisioning together: bbv's switch touched all three in one 26-commit pull request.
What happened
A 26-commit pull request merged on February 12, 2026 made the switch: tracing moved to OpenTelemetry and Langfuse spans, the dataset API moved from Phoenix to Langfuse, and Langfuse is provisioned automatically.
“Migrate tracing to use OTEL & Langfuse spans…”
mfundn
- Public engineering write-up (GitHub pull request)
- Code change merged
Deployment details
bbv's AI Hub core, with agent tracing, datasets and automatic provisioning of the trace backend.
- Public engineering write-up (GitHub pull request)
- Code change merged
Team
bbv AI Hub core
Duration
Merged February 12, 2026
Workload
Agent tracing and evaluation datasets
Product version
Langfuse through OpenTelemetry spans, replacing Arize Phoenix
In SignalsCounted once, as the contributor's own pull request. In the record

MedicSep 23, 2026GitHub (medic/cht-agent)Medic moved its health-worker agent's Langfuse tracing to the OpenTelemetry-based JS SDK v5 ahead of the Cloud cutover.
alexosugoMaintainer, medic/cht-agentMedic
The Community Health Toolkit agent, on Langfuse Cloud.
GitHub (medic/cht-agent), published Sep 23, 2026Read the write-up (opens GitHub (medic/cht-agent) in a new tab)
Observation
Medic moved its health-worker agent's Langfuse tracing to the OpenTelemetry-based JS SDK v5 ahead of the Cloud cutover.
Takeaway
Start new Langfuse integrations on the v4-era SDKs; tracing added on the v3 package in September 2026 had to be migrated within weeks.
What happened
A pull request opened September 23, 2026 and merged October 1 migrated the agent's new Langfuse tracing from the legacy v3 npm package to the OpenTelemetry-based JS SDK v5, because Langfuse Cloud removes legacy batch ingestion on November 16, 2026.
“…migrates the Langfuse tracing added there from the legacy langfuse v3 npm package to the OpenTelemetry-based JS SDK v5…”
alexosugo
Public engineering write-up (GitHub pull request)
Code change merged
Deployment details
The Community Health Toolkit agent, on Langfuse Cloud.
Public engineering write-up (GitHub pull request)
Code change merged
Team
Medic cht-agent maintainers
Workload
Community Health Toolkit agent tracing
Product version
Langfuse JS SDK v5, replacing the v3 npm package
In SignalsCounted once, as the maintainer's own pull request. In the record
ApifySep 14, 2026GitHub (apify/apify-mcp-server)Apify replaced its Arize Phoenix eval runner with Langfuse-based gates, because the old runner's verdicts did not carry over to real use.
Jiří SpilkaEngineerApify
Evals that check tool selection for Apify's MCP server.
GitHub (apify/apify-mcp-server), published Sep 14, 2026Read the write-up (opens GitHub (apify/apify-mcp-server) in a new tab)
Observation
Apify replaced its Arize Phoenix eval runner with Langfuse-based gates, because the old runner's verdicts did not carry over to real use.
Takeaway
Judge an eval harness by whether its verdicts hold in the client your users run; Apify moved its gates to Langfuse once the old loop stopped matching Claude Code.
What happened
In a pull request opened September 14, 2026 and merged September 17, Apify deleted the Phoenix eval runner from apify-mcp-server after Langfuse-based two-tier gates replaced it. The runner's custom loop could not represent built-in tool competition or Claude Code's prompting, so its verdicts did not transfer.
Public engineering write-up (GitHub pull request)
Code change merged
Deployment details
Evals that check tool selection for Apify's MCP server.
Public engineering write-up (GitHub pull request)
Code change merged
Team
Apify MCP server team
Workload
Tool-selection evals for an MCP server
Product version
Langfuse-based two-tier eval gates
In SignalsCounted once, as the engineer's own pull request. In the record
createsomethingtodayJul 6, 2026GitHub (createsomethingtoday/create-something-monorepo)CREATE SOMETHING retired Braintrust for Langfuse across its evals, telemetry and CI workflows in one pull request.
createsomethingtodayRepository ownerCREATE SOMETHING
The CREATE SOMETHING monorepo: evals, telemetry, prompts and MCP services.
GitHub (createsomethingtoday/create-something-monorepo), published Jul 6, 2026Read the write-up (opens GitHub (createsomethingtoday/create-something-monorepo) in a new tab)
Observation
CREATE SOMETHING retired Braintrust for Langfuse across its evals, telemetry and CI workflows in one pull request.
Takeaway
A full move between eval platforms fits in one reviewed pull request when evals, telemetry and CI each sit behind their own module; rerun every suite before deleting the old one.
What happened
A pull request merged on July 6, 2026 replaced Braintrust's eval, telemetry, package, script, environment and workflow surfaces with Langfuse equivalents, then reran nine eval suites against Langfuse. The pull request does not say why the team moved.
Public practitioner write-up (GitHub pull request)
Code change merged
Deployment details
The CREATE SOMETHING monorepo: evals, telemetry, prompts and MCP services.
Public practitioner write-up (GitHub pull request)
Code change merged
Team
CREATE SOMETHING
Workload
Evals, telemetry, scripts and CI workflows
Product version
Langfuse, replacing Braintrust
Measurement
Nine eval suites rerun against Langfuse
In SignalsCounted once, as the repository owner's own pull request. In the record

Evaluated it1

AmitAuthorNonBioS
NonBioS, an AI engineering agent product, after rewriting its systems to drop LangChain.
NonBioS blog, published Mar 5, 2025Read the write-up (opens NonBioS blog in a new tab)
Observation
NonBioS evaluated about a dozen tracing tools and chose Langfuse over Helicone, after ruling out LangSmith for its LangChain coupling.
Takeaway
Shortlist on how traces group into sessions in your product, and test that first; it decided NonBioS's choice.
What happened
NonBioS listed about a dozen offerings, including LangSmith, Lunary, Arize Phoenix, Portkey, Datadog, Langfuse and Helicone (NonBioS blog, Mar 5, 2025). It ruled out LangSmith as strongly coupled to LangChain, looked at Phoenix only briefly, and narrowed to Langfuse and Helicone. It chose Langfuse for session tracking, open source it can extend, and prompt version control.
“Session tracking capabilities significantly impacted our decision. Langfuse offers a tracing feature that connects sequences of LLM calls to single sessions. This aligned well with our needs and appeared straightforward in initial testing.”
Amit
- Public organization write-up
- No team metrics reported
Deployment details
NonBioS, an AI engineering agent product, after rewriting its systems to drop LangChain.
- Public organization write-up
- No team metrics reported
Team
NonBioS engineering
Workload
Tracing an AI engineering agent's LLM calls by session
In SignalsCounted once, as NonBioS's own post. In the record

Adopted it10

RampSep 17, 2026ClickHouse on YouTubeRamp's Reflect agent scores and clusters traces in self-hosted Langfuse and proposes fixes that owners accept.
Ryan and DavidRamp engineersRamp
Agents across expense, close and an internal coding agent, on self-hosted Langfuse with ClickHouse Cloud.
ClickHouse on YouTube, published Sep 17, 2026Watch the talk (opens ClickHouse on YouTube in a new tab)
Observation
Ramp's Reflect agent scores and clusters traces in self-hosted Langfuse and proposes fixes that owners accept.
Takeaway
Once traces are complete, an agent that reads them can propose the fixes; keep a human owner accepting each one.
What happened
Ramp runs its agent tracing on self-hosted Langfuse. Its Reflect agent scores and clusters those traces and proposes fixes for an owner to accept. Reflect covers 12 agents and more than 120 merged suggestions; on the Research agent alone, 60 or more merged suggestions cut tokens 10 to 20%, tool calls 15% and session time 30% (talk on ClickHouse's YouTube channel, Sep 17, 2026).
Vendor customer story (talk on the YouTube channel of ClickHouse, which owns Langfuse)
Token and latency changes reported
Deployment details
Agents across expense, close and an internal coding agent, on self-hosted Langfuse with ClickHouse Cloud.
Vendor customer story (talk on the YouTube channel of ClickHouse, which owns Langfuse)
Token and latency changes reported
Team
Ramp engineering and data platform
Workload
12 production agents
Product version
Self-hosted Langfuse on S3, Redis, Postgres, ECS and ClickHouse Cloud
Measurement
Tokens, tool calls and session time before and after accepted fixes
In SignalsNot counted. ClickHouse, which owns Langfuse, published the talk, so its record is listed and counts toward no figure.
MerckSep 2, 2025Langfuse customer storyMerck's central AI platform team provisions self-hosted Langfuse to its GenAI teams through an automated request portal.
Harsha GurulingappaGlobal Head of AI & Machine Learning PracticeMerck
Central AI & ML platform for Healthcare, Life Science, and Electronics, with more than 200 regular users.
Langfuse customer story, published Sep 2, 2025Read the story (opens Langfuse customer story in a new tab)
Observation
Merck's central AI platform team provisions self-hosted Langfuse to its GenAI teams through an automated request portal.
Takeaway
At company scale, provision tracing through APIs from a request portal, and start the shared evaluation dataset once every team traces.
What happened
Merck's AI & Machine Learning team runs shared platforms across its business sectors. Teams document use cases in a portal and request Langfuse. After approval, API-driven pipelines create projects and users, plus API keys. Pharmaceutical data stays in Merck's infrastructure. About 80 of more than 300 GenAI use cases run on Langfuse, mostly RAG apps (Langfuse customer story, Sep 2, 2025). The team plans a company-wide benchmark dataset from vetted production outcomes.
Vendor customer story (Langfuse site)
Adoption counts reported
Deployment details
Central AI & ML platform for Healthcare, Life Science, and Electronics, with more than 200 regular users.
Vendor customer story (Langfuse site)
Adoption counts reported
Team
Merck's AI & Machine Learning team, serving all three business sectors
Workload
GenAI use cases, mostly RAG applications
Product version
Self-hosted Langfuse provisioned through its APIs
Measurement
Use cases and people on the platform
In SignalsNot counted. Langfuse published this story on its own site, so GILD counts it nowhere in the Signals.
SumUpAug 11, 2025Langfuse customer storySumUp debugs its AI merchant support asynchronously in Langfuse, and reports half of conversations resolved without a human.
Ana CasadoHead of Operations Data and AISumUp
AI support for SumUp's merchants, built with its machine learning team.
Langfuse customer story, published Aug 11, 2025Read the story (opens Langfuse customer story in a new tab)
Observation
SumUp debugs its AI merchant support asynchronously in Langfuse, and reports half of conversations resolved without a human.
Takeaway
Track the deflection rate beside the traces, so the support team can see which failures cost a human handoff.
What happened
SumUp uses Langfuse for observability and asynchronous debugging of its AI merchant support. SumUp reports a 50% deflection rate, about 300,000 merchant requests a month, or 10,000 a day, that no longer reach human support (Langfuse customer story, Aug 11, 2025).
Vendor customer story (Langfuse site)
Deflection rate reported
Deployment details
AI support for SumUp's merchants, built with its machine learning team.
Vendor customer story (Langfuse site)
Deflection rate reported
Team
SumUp merchant support AI, operations data and machine learning
Workload
AI merchant support, about 300,000 requests a month
Measurement
Share of conversations resolved without a human
In SignalsNot counted. Langfuse published this story on its own site, so GILD counts it nowhere in the Signals.
PinterestJul 20, 2026AI Engineer World's Fair 2026Pinterest's Medic team reads each agent run in Langfuse as a waterfall of steps to find what caused a weak diagnosis.
Drasko ProfirovicStaff EngineerPinterest
Medic, an agent that diagnoses failing Apache Spark jobs through MCP tools over job logs and metrics.
AI Engineer World's Fair 2026, published Jul 20, 2026Watch the talk (opens AI Engineer World's Fair 2026 in a new tab)
Observation
Pinterest's Medic team reads each agent run in Langfuse as a waterfall of steps to find what caused a weak diagnosis.
Takeaway
Trace an agent from its first production run, then read failures step by step before changing prompts, and keep a replay harness to grade each change.
What happened
To improve Medic, the team invested in observability and testability. It publishes the agent's OpenTelemetry traces to Langfuse and reads each run as a waterfall of steps to find the cause of lower-quality answers. Prompt changes are then graded with offline evals in the team's own replay harness (AI Engineer World's Fair talk, July 20, 2026).
“To improve the system, we invested in observability and testability. We used OpenTelemetry to publish traces to Langfuse, and by viewing the agent's execution as a waterfall diagram of steps, we could better understand the cause of lower quality responses.”
Drasko Profirovic
Conference talk
No team metrics reported
Deployment details
Medic, an agent that diagnoses failing Apache Spark jobs through MCP tools over job logs and metrics.
Conference talk
No team metrics reported
Team
Pinterest data platform team behind Medic
Workload
Agent that diagnoses failing Apache Spark jobs
Product version
Langfuse fed by OpenTelemetry
In SignalsCounted once, as the speaker's own talk. In the record
CrestaMar 26, 2026Cresta engineering blogCresta self-hosts one Langfuse instance per Kubernetes cluster and one Langfuse organization per customer, so conversation traces never leave the cluster that produced them.
James IdzikSoftware EngineerCresta
AI agents for customer conversations. Go and Python services. Self-hosted Langfuse in production.
Cresta engineering blog, published Mar 26, 2026Read the write-up (opens Cresta engineering blog in a new tab)
Observation
Cresta self-hosts one Langfuse instance per Kubernetes cluster and one Langfuse organization per customer, so conversation traces never leave the cluster that produced them.
Takeaway
If traces hold customer conversations, decide isolation and retention before you instrument: per-customer projects or organizations, redaction before export, and an automatic delete date.
What happened
Cresta runs intent detection and retrieval, plus tool calls and safety checks, across Go and Python services with separate tool functions. Trace context travels in gRPC metadata, so Langfuse reconstructs one tree per customer turn. Each customer has a separate Langfuse organization and credentials resolved from request context. Cresta redacts PII before recording and deletes traces after four weeks (Cresta engineering blog, Mar 26, 2026). Engineers inspect the tree to find the failed step, including misread intent that misdirects later steps.
“We self-host Langfuse rather than using a managed SaaS offering.”
James Idzik
Public engineering write-up (co-written with Langfuse)
No team metrics reported
Deployment details
AI agents for customer conversations. Go and Python services. Self-hosted Langfuse in production.
Public engineering write-up (co-written with Langfuse)
No team metrics reported
Team
Cresta engineering
Workload
Multi-service agent pipelines for customer conversations
Product version
Self-hosted Langfuse, one instance per Kubernetes cluster, one organization per customer
In SignalsCounted once, as Cresta's own engineering post. In the record
Khan AcademyFeb 5, 2025AI Engineer World's Fair 2024Khan Academy's AI platform team gives every conforming Khanmigo service Langfuse tracing through one Go interface.
Shawn JanseparDirector of AI and Learning (self-described in the talk)Khan Academy
Khanmigo AI tutor, where an AI platform team owns the developer experience for product teams.
AI Engineer World's Fair 2024, published Feb 5, 2025Watch the talk (opens AI Engineer World's Fair 2024 in a new tab)
Observation
Khan Academy's AI platform team gives every conforming Khanmigo service Langfuse tracing through one Go interface.
Takeaway
Make tracing a property of the platform interface, so product teams get it by conforming and never wire Langfuse themselves.
What happened
The AI platform team built a component interface in Go: a service that conforms to it gets tracing and Langfuse without extra work. All stack traces are now available in Langfuse, so engineers open a failing case there and debug from it (AI Engineer World's Fair 2024 talk, published Feb 5, 2025).
“…we recently built this component interface in Go that when you as a developer conform to that component interface, it automatically think-- means that things like tracing and Langfuse are available as long as you conform.”
Shawn Jansepar
Conference talk
No team metrics reported
Deployment details
Khanmigo AI tutor, where an AI platform team owns the developer experience for product teams.
Conference talk
No team metrics reported
Team
Khanmigo AI platform team
Workload
Khanmigo AI tutor services
Product version
Langfuse through a shared Go component interface
In SignalsCounted once, as the speaker's own talk. In the record
SuperDialApr 18, 2025AI Engineer Summit 2025SuperDial self-hosts Langfuse for all logging and observability because its voice agents place healthcare calls.
Nik CaryotakisStaff Software EngineerSuperDial
Voice agents that place long healthcare phone calls, with a TensorZero gateway in front of the models.
AI Engineer Summit 2025, published Apr 18, 2025Watch the talk (opens AI Engineer Summit 2025 in a new tab)
Observation
SuperDial self-hosts Langfuse for all logging and observability because its voice agents place healthcare calls.
Takeaway
When calls or chats carry regulated data, self-host the tracer with the rest of the stack, and decide that before the first call is recorded.
What happened
SuperDial runs all logging and observability for its voice agents on self-hosted Langfuse, and self-hosts its other tooling for the same reason: the calls carry healthcare data. It uses Langfuse for anomaly detection and evals (AI Engineer Summit 2025 talk, published Apr 18, 2025).
“And then all of our logging and observability, we self-host Langfuse, and we self-host these things also because these are like healthcare calls.”
Nik Caryotakis
Conference talk
No team metrics reported
Deployment details
Voice agents that place long healthcare phone calls, with a TensorZero gateway in front of the models.
Conference talk
No team metrics reported
Team
SuperDial voice AI engineering
Workload
Healthcare phone calls placed by voice agents
Product version
Self-hosted Langfuse
In SignalsCounted once, as the speaker's own talk. In the record
AmplifonApr 10, 2026AI Engineer Europe 2026Amplifon deployed Langfuse at the platform level of its AmplifAI program, so every team that starts from its agent blueprints traces and evaluates from day one.
Sonny MerlaGlobal Data Science & AI ManagerAmplifon
AmplifAI, Amplifon's agent program, with MCP and A2A server blueprints built with Quantyca engineers.
AI Engineer Europe 2026, published Apr 10, 2026Watch the talk (opens AI Engineer Europe 2026 in a new tab)
Observation
Amplifon deployed Langfuse at the platform level of its AmplifAI program, so every team that starts from its agent blueprints traces and evaluates from day one.
Takeaway
Put the tracer in the template every team starts from, so tracing and evals are present before the first agent ships.
What happened
Amplifon's MCP and A2A blueprints ship with Dockerfiles, authentication and cost tracking, and with an integration to Langfuse deployed at the platform level. Development teams that start from a blueprint can trace their agents, run evaluations and check how each agent performs, whatever framework they use (AI Engineer Europe 2026 talk, Apr 10, 2026).
Conference talk
No team metrics reported
Deployment details
AmplifAI, Amplifon's agent program, with MCP and A2A server blueprints built with Quantyca engineers.
Conference talk
No team metrics reported
Team
Amplifon AI Center of Excellence, with Quantyca
Workload
MCP and A2A agent servers built from shared blueprints
Product version
Langfuse deployed at the platform level
In SignalsCounted once, as the speakers' own talk. In the record
DKBAug 31, 2026Langfuse customer storyDKB screens its banking assistant's customer conversations in self-hosted Langfuse, and says it would have to stop its agents if the screening went off.
David HinrichsSenior Machine Learning ScientistDKB
In-app assistant for about 6 million customers (Langfuse customer story, Aug 31, 2026). 10,000 to 20,000 conversations a day. Regulated bank.
Langfuse customer story, published Aug 31, 2026Read the story (opens Langfuse customer story in a new tab)
Observation
DKB screens its banking assistant's customer conversations in self-hosted Langfuse, and says it would have to stop its agents if the screening went off.
Takeaway
In a regulated product, treat screening and review of traces as part of the license to operate, and design who may see which conversations before you route them anywhere.
What happened
DKB required a self-hosted open-source tool before sending customer data, with access for service experts as well as engineers. Langfuse traces every interaction. Automated checks flag investment advice and ungrounded answers, plus unhappy customers. Specialists review flagged sessions by service area and severity. DKB uses two production instances because colleagues' and minors' conversations cannot be screened automatically. Flagged behavior becomes datasets for regression tests before release.
Vendor customer story (Langfuse site)
Volume reported
Deployment details
In-app assistant for about 6 million customers (Langfuse customer story, Aug 31, 2026). 10,000 to 20,000 conversations a day. Regulated bank.
Vendor customer story (Langfuse site)
Volume reported
Team
DKB machine learning team with customer service and banking specialists
Duration
About a year
Workload
Agentic in-app banking assistant, 10,000 to 20,000 conversations a day
Product version
Self-hosted Langfuse, two production instances
In SignalsNot counted. Langfuse published this story on its own site, so GILD counts it nowhere in the Signals.
RavennaJul 24, 2026Langfuse customer storyRavenna links every AI interaction to a Langfuse trace, so customer-success staff diagnose support-agent failures without pulling in engineers.
Taylor HallidayCo-founder and CEORavenna
Support agents that handle hundreds of systems and request types across customer companies.
Langfuse customer story, published Jul 24, 2026Read the story (opens Langfuse customer story in a new tab)
Observation
Ravenna links every AI interaction to a Langfuse trace, so customer-success staff diagnose support-agent failures without pulling in engineers.
Takeaway
Give support staff trace access and a handoff format, so engineers see only the failures a trace cannot explain.
What happened
Ravenna uses Langfuse to tie each AI interaction to a trace. Its customer-success team triages and diagnoses issues on its own: it spots the problem in the trace, documents it, and hands it off with full context (Langfuse customer story, Jul 24, 2026).
Vendor customer story (Langfuse site)
No team metrics reported
Deployment details
Support agents that handle hundreds of systems and request types across customer companies.
Vendor customer story (Langfuse site)
No team metrics reported
Team
Ravenna engineering and customer success
Workload
AI support agents across customer systems
In SignalsNot counted. Langfuse published this story on its own site, so GILD counts it nowhere in the Signals.

Left it or turned it down2

Bayes Impact is leaving Langfuse: a merged pull request has its platform emit plain OpenTelemetry, and a collector sends LLM traces to Arize Phoenix.

digithings.ai approved self-hosted Langfuse to replace LangSmith, then chose Cloudflare-native tracing with MLflow instead.

Switched to or from it4

bbv's AI Hub replaced Arize Phoenix with Langfuse in February 2026, moving its tracing, datasets and trace-backend provisioning.

Medic moved its health-worker agent's Langfuse tracing to the OpenTelemetry-based JS SDK v5 ahead of the Cloud cutover.

Apify replaced its Arize Phoenix eval runner with Langfuse-based gates, because the old runner's verdicts did not carry over to real use.

CREATE SOMETHING retired Braintrust for Langfuse across its evals, telemetry and CI workflows in one pull request.

Evaluated it1

NonBioS evaluated about a dozen tracing tools and chose Langfuse over Helicone, after ruling out LangSmith for its LangChain coupling.

Adopted it10

Ramp's Reflect agent scores and clusters traces in self-hosted Langfuse and proposes fixes that owners accept.

Merck's central AI platform team provisions self-hosted Langfuse to its GenAI teams through an automated request portal.

SumUp debugs its AI merchant support asynchronously in Langfuse, and reports half of conversations resolved without a human.

Pinterest's Medic team reads each agent run in Langfuse as a waterfall of steps to find what caused a weak diagnosis.

Cresta self-hosts one Langfuse instance per Kubernetes cluster and one Langfuse organization per customer, so conversation traces never leave the cluster that produced them.

Khan Academy's AI platform team gives every conforming Khanmigo service Langfuse tracing through one Go interface.

SuperDial self-hosts Langfuse for all logging and observability because its voice agents place healthcare calls.

Amplifon deployed Langfuse at the platform level of its AmplifAI program, so every team that starts from its agent blueprints traces and evaluates from day one.

DKB screens its banking assistant's customer conversations in self-hosted Langfuse, and says it would have to stop its agents if the screening went off.

Ravenna links every AI interaction to a Langfuse trace, so customer-success staff diagnose support-agent failures without pulling in engineers.