← Back to News

2026-06-15

Hopsworks 5.0 - Where AI Agents Go to Work

Hopsworks 5.0 - Where AI Agents Go to Work

Hopsworks 5.0 is focussed on improving developer productivity in Data and AI. With native coding agent support, developers can build production AI systems (batch, real-time, and agentic) and dashboards (Apache Superset) in one tenth of the time it took previously*. 

Hopsworks for Agents

Version 5.0 brings first-class, end-to-end support for AI and agentic workloads. Four new capabilities make it real.

Hopsworks Terminal

A browser-based terminal backed by a dedicated container, with credentials for coding agents like Claude Code and OpenAI Codex mounted directly from HopsFS. That means an agent running inside the Terminal can drive the platform on your behalf; securely, and out of the box. Per-session heartbeats keep long-running work alive with clear timeout notices, so your agents never silently disappear on idle.

Agent Workflows

Headless agents can be run as Hopsworks jobs. Choose a provider per job (Anthropic, OpenAI, Open-Weight Models, etc), pass account-level and per-job environment variables, and use the same Hopsworks job scheduler that already runs your pipelines. 

Agent Deployments

Take a local script or a pyproject.toml-rooted package and ship it as a server-only deployment, powered by our new python-agent-pipeline base image. Prototype to production endpoint, in one clean move. Full OpenTelemetry tracing for agent deployments; online, offline, or both; behind a single cluster-wide toggle. Finally, you can see exactly what your agents did, with the observability your production workloads deserve.

Analytics with Apache Superset and Trino

Trino and Apache Superset are now multi-tenant services in Hopsworks. You can securely share Trino and Apache Superset among your teams, and your end users can now build and deploy dashboards from federated data sources in seconds using coding agents.

A platform that guides you

Great tools don't just give you power; they help you wield it. Hopsworks 5.0 introduces a layer of platform intelligence that turns hard-won institutional knowledge into guided, in-product steps. Scheduling a job now means smart suggestions for the right compute, arguments, and data sources; so the decisions that used to live in someone's head live in the platform instead.

Platform Instrumentalisation

The Hopsworks CLI and its companion skills expose the Feature Store and the wider platform as a direct, first-class tool interface for LLMs. Any AI agent can drive an entire Hopsworks cluster through them; and they sit at the center of the 5.0 release.

Coverage is complete. The full Platform and FTI lifecycle; feature groups, feature views, training datasets, transformations, tags, and lineage. The full job lifecycle; create, run, stop, status, logs, history, schedule, schedule-info, unschedule, and delete; across Python, PySpark, Spark, and Ray. And a dedicated hops agent command group that mirrors hops deployment, so working with agents feels exactly like working with everything else.

More Power to the Feature Store

The Feature Store that thousands of teams rely on gets sharper in 5.0:

  • Transformation chaining lets you compose transformation functions into an execution DAG; with topological ordering, duplicate-output detection, and inline reporting of cycles and unknown inputs. The resolved DAG is rendered right on your Feature View and Feature Group overview pages.
  • Point-in-time lookback windows bound how far your joins reach back, for tighter correctness and far less wasted scan.
  • Extra batch filters let you refine batch reads from a feature view without ever reshaping the view itself.
  • Online table indexing brings secondary indexes and configurable primary-key index types to online feature groups, so your reads match your access patterns.
  • Native Apache Iceberg support, across the platform.

Connect more

As your data - and your models - live in more places than ever, with Hopsworks 5.0, we extended DLThub support and now include resource estimation for DLThub jobs. MongoDB joins as a native data source. The Databricks Unity Catalog connector integrates natively with OAuth M2M authentication. You can create data sources directly through the API. And you can now pull models straight from Hugging Face into Hopsworks. 

Analytics and SQL, now first-class

Interactive analytics is now supported fully in 5.0, SQL becomes a first-class citizen on the platform, powered by Trino as a query engine and Superset dashboards. Query your features, explore your data, and build dashboards without leaving the platform or wiring up a separate BI stack.

A faster UI, a better UX. 

Hopsworks 5.0 is a ground-up redesign and reimplementation of our interface, migrating off our bespoke Quartz onto shadcn. The result is a platform that's faster across both self-managed and SaaS deployments, and more alive in the moment; with dynamic component sizing and layout that adapts to new tooling like the Terminal, plus Superset dashboards embedded directly in the platform.

Upgraded Software

  • Airflow → 3.0.6
  • Default Python → 3.12
  • TensorFlow → 2.19.0
  • scikit-learn → 1.8.0
  • RonDB → 26.2.10
  • OpenSearch → 2.19.5
  • vLLM → multi-version support

Get started today

Hopsworks 5.0 is generally available now for SaaS and self-managed customers. Open the Terminal, bring your agents, and build your first FTI pipeline without ever leaving the platform.