Langfuse roadmap
Langfuse is open source and we want to be fully transparent what we're working on and what's next. This roadmap is a living document and we'll update it as we make progress.
Your feedback is highly appreciated. Feel like something is missing? Add new ideas on GitHub or vote on existing ones. Both are a great way to contribute to Langfuse and help us understand what is important to you.
Vision and direction
Langfuse should become the open data and evaluation layer that helps humans, and eventually agents, improve agents. We optimize for one product loop above all else: track, understand, evaluate, and improve agentic systems.
The strategic choice is to stay neutral in the execution layer. Langfuse should not become an opinionated agent framework or runtime. Instead, Langfuse should own the improvement loop around agentic software: understand agent behavior, segment it into useful views, turn production failures into datasets, run experiments, and automate repeated workflows through APIs, the CLI, skills, and an in-product agent.
The long-term direction is auto-optimizing agents: connect tracing and your code repository, and Langfuse can manage the agent improvement loop for you. Langfuse understands the instructions, prompts, evals, and skill files that define your system; manages versions; runs evaluations; proposes or triggers experiments; and keeps humans involved for the highest-leverage judgments.
Active development
The Q2 2026 focus is to make the existing foundation excellent and connect the pieces into a continuous improvement loop for agents.
Agent observability and views
- Make the v4 observations table, filter sidebar, saved views, and default views excellent for agent traces.
- Build agent-level views for traces per agent, cost, latency, steps, tool calls, and aggregate step/tool behavior.
- Improve trace detail pages for long-running agent traces, including compact representations, selected JSON paths, and better ways to move from charts to the underlying spans.
- Improve full-text search, metadata filtering, custom dimensions, and dashboard-to-trace workflows so teams can slice observations with less noise.
Evals and experiments
- Ship public APIs for experiments and evaluators.
- Scale the evaluator data model and support new evaluator types.
- Improve experiment charts, comparison flows, evaluator management, and the evaluator template library.
- Expand code-based evals, categorical and boolean judges, free-text scores, multimodal datasets, and the trace-level eval deprecation path.
Workflow automation and agents
- Build the first in-product Langfuse agent for reading Langfuse data, using screen context, and helping with tasks such as comparing traces.
- Use skills, guides, and academy content to automate AI engineering workflows outside the product before packaging the best ones in-product.
- Improve the Langfuse CLI, MCP surfaces, and skill management so external agents can inspect data shape, query Langfuse efficiently, and execute common workflows.
- Prioritize repeatable workflows such as low-score analysis, failure clustering, evaluator setup, production-to-dataset refreshes, synthetic data generation, and experiment triggering.
Platform reliability and scale
- Finish the v4 rollout across Langfuse Cloud and self-hosted deployments.
- Continue scaling ingestion for large agent workloads and make read paths faster through pre-aggregation where needed.
- Fix event-loop and public API reliability issues, improve queue reliability, and make SLOs actionable across core product areas.
- Make system integration points such as blob exports, S3 exports, public APIs, metrics, observations access, and the CLI boringly reliable.
Alerts, workflows, and enterprise controls
- Ship alerting for evals, metrics, and operational thresholds across delivery channels such as Slack, PagerDuty, webhooks, and email.
- Explore webhooks and automations for observability and evaluation events.
- Improve API-key scoping, move toward bearer keys, and expand admin controls for enterprise deployments.
- Improve the self-hosted and Helm chart experience, and explore hybrid or BYOC deployment models for customers that need stronger data isolation or direct ClickHouse access.
Multimodal and playground
- Close multimodal gaps so traces, playground, datasets, and evals feel like one consistent system.
- Make the playground more stateful and collaborative so teams can invest in reusable debugging and experimentation setups.
12-month product direction
Views as the platform primitive
Views should become the primitive for slicing observations into useful product surfaces. A view defines which observations matter, how they are grouped, which attributes are shown, which metrics and scores matter, and which downstream actions are available. This unlocks agent overview dashboards, default templates, semantic clustering, evaluation distribution comparisons, and workflow triggers.
Preference layer
Human judgment remains the ground truth for evaluating agents. Langfuse should make it easier to capture explicit feedback, derive implicit signals, align LLM-as-a-judge evaluators with human preferences, and route low-confidence cases back into human review.
Semantic grouping
As agents move from routed sub-agent systems to broader dynamic agents, fixed labels are not enough. Langfuse should help teams discover meaningful interaction groups within a filtered view, compare scores across those groups, and turn recurring failures into datasets or experiments.
Experiments as the hill-climbing surface
Experiments should become a flagship workflow for comparing prompt, model, and runtime changes. Langfuse should make baselines, run comparisons, annotations, metrics, and next actions easy enough that teams naturally use experiments as their agent improvement loop.
Managed improvement loop
The end state is that Langfuse can monitor an agent system, propose or run experiments, refresh test sets from production, assign annotation work when human input is needed, and report how the system is improving over time.
🚀 Recently released
10 most recent changelog items:
- Self-Service Enterprise SSO Setup
- Experiments CI/CD integration
- Langfuse Cloud Japan
- Amazon Bedrock API Keys
- Experiments as a First-Class Concept
- Free-Form Text Scores
- Boolean LLM-as-a-Judge Scores
- Updates to Dashboards
- Categorical LLM-as-a-Judge Scores
- Simplify Langfuse for Scale
Subscribe to our mailing list to get occasional email updates about new features.
🙏 Feature requests and bug reports
The best way to support Langfuse is to share your feedback, report bugs, and upvote on ideas suggested by others.
