Release Notes¶

A history of Agentlab-specific changes on top of Langflow, newest first. Internal build, CI and rebase-maintenance work is omitted.

v1.3.1313 Oct 2026

How Agentlab versions work

Agentlab tracks its own version line, independent of Langflow. A minor release (e.g. 1.2.x → 1.3.x) marks a rebase onto a newer upstream Langflow release, bringing in all of that version's improvements. Patch releases add Agentlab features and fixes on top of the current base.

Each group below notes the underlying Langflow version. For upstream changes, see the Langflow release notes.

1.3.x¶

1.3.13¶

New Extraction progress for KARLI documents — the Read File component now reports progress while the KARLI extraction backend processes a document: pages done out of the total, updated each time a section of the document is finished. The progress appears in the component's logs and is sent on the flow's run stream, from where Karli Studio shows it as a progress bar. Progress is currently reported only with the karli/data-extraction-moe-latest extraction model; other models extract as before, without progress. Long documents no longer fail after five minutes.
New Files Table output on Read File — with Allow User File Upload switched on, the Read File component offers a third output, Files Table: one row per uploaded file with its path, name, extracted content, error, extraction task ID and model, in the same columns whatever the extraction backend. Later components can then handle the files one by one, for example to delete them after use. When no file has been uploaded, the table is empty instead of raising an error.

1.3.12¶

Fixed Agent tools with Langfuse tracing — Agents on instances with Langfuse tracing enabled call their tools, such as web search, reliably again, and those tool calls show up in Langfuse traces as usual. Previously, tool calls failed on these instances. No configuration change is needed.

1.3.11¶

Fixed Default model of new Agents — New Agent components now come preselected with karli/KARLI-MoE-latest on the KARLI provider, ready to run. Previously, the preselected model was not available, so a new Agent's first run failed. Agents added under 1.3.10 are updated automatically the next time their flow is opened in the editor; flows that only run through the API keep their previous selection until then.
Fixed Model selection with shared model names — When two providers offer a model under the same name, the model picker now keeps them apart, so each selection is saved and shown under the provider you chose. Previously, a selection could end up under the other provider. When a flow is opened, a saved selection that no provider offers is replaced with an available model.
Fixed Agent tools with Prometheus metrics — Agents on instances with Prometheus metrics enabled call their tools reliably again, and token metrics keep being recorded. Previously, tool calls failed on these instances. No configuration change is needed.

1.3.10¶

Fixed Builder assistant visibility — on an instance with the builder assistant switched off (LANGFLOW_AGENTIC_EXPERIENCE), the assistant no longer appears in the interface. Its button, keyboard shortcut and prompt box were still offered and simply failed when used. The flow templates on the new-flow screen are unaffected and stay available.
Changed Builder assistant marked as Beta — the assistant now carries a Beta label, reflecting that it is still a preview feature.

1.3.9¶

Changed Component logging — log output written by a component at info, warning or error is no longer emitted on an instance running at the usual log level. This guards against custom components writing sensitive data, such as conversation content, into the central log store. Component diagnostics belong at debug, which is off in production: write them there and run the instance at LANGFLOW_LOG_LEVEL=DEBUG when you need to read them. Errors raised by a component are reported by Agentlab itself and appear as before.
Fixed KARLI (AISL) models — the provider listed models the gateway does not serve, so selecting the pre-selected default failed with "Model not available". The model list now comes from the gateway itself. Model Providers →
Improved KARLI (AISL) model list — chat, embedding and extraction models are read from the gateway when the instance starts and refreshed hourly, so models added on the gateway appear without an Agentlab release. If the gateway is unreachable the previously known list is kept, never an empty one.

1.3.8¶

Fixed Builder assistant — enabling the builder assistant (LANGFLOW_AGENTIC_EXPERIENCE) no longer breaks sign-in. A constant the agentic setup needs had been lost in an earlier rebase, so logging in failed outright on any instance with the feature switched on.

1.3.7¶

New KARLI (AISL) model provider — a new provider reaching chat, embedding and document-extraction models through the AI Service Layer gateway. It needs a single API key instead of the two base URLs and two keys the existing KARLI provider requires; the gateway address is an instance setting. The existing KARLI provider is unchanged and stays available — both can run side by side. Model Providers →
New AISL document extraction — the Read File component gains an aisl extraction backend. Available extraction services are read live from the gateway, so newly added ones appear without an update. Requires object storage (LANGFLOW_STORAGE_TYPE=s3), because the gateway fetches the document from a presigned URL. Document Extraction →
Fixed Provider restriction — the MODEL_PROVIDERS environment variable takes effect again. The setting had silently stopped being applied during an earlier rebase, so instances offered every configured provider regardless of the list. Model Providers →

1.3.6¶

New LLM usage metrics — instances can now report token counts, call counts and response times per model and per flow, on a dedicated metrics port. Usage is grouped by flow name, so you can see which flows are consuming what. This works independently of Langfuse, so instances without tracing configured still report usage. Observability →
Fixed Metrics endpoint — the metrics port returned only generic process statistics and never reported any Langflow activity. File-upload metrics now appear as well.

1.3.5¶

New Agentlab version — the account menu now shows the Agentlab release version alongside the underlying Langflow version, so you can tell exactly which build you're running.

1.3.4¶

Improved Structured Output — the Format Instructions field is now available in the component's settings, so you can customize how the model formats its extracted output. It was previously hidden from view.
Fixed Embedding Model — a newly added Embedding Model component works on the first run. Previously it failed with a "no embedding class" error until you re-selected the model from the dropdown.
New Embedding overrides — optional Model Name and Provider inputs let a global variable choose the embedding model at runtime.
Fixed API Keys — deleting an API key works again; it previously returned an error.

1.3.3¶

Fixed Tool Calls — custom tool titles and display modes (introduced in 1.2.7) are saved again. After the 1.10.3 rebase they were reset when the Actions sidebar was reopened; the values now persist. Tool Calls →

1.3.2¶

Improved Observability — trace span labels and types are normalized to a consistent lowercase form, so spans group and filter reliably across runs.

1.3.0¶

Changed Rebased onto Langflow 1.10.3 — the full set of upstream improvements comes with it.
Improved KARLI embedding catalog — refreshed, with the docs updated to match.
Changed Context Input — adapted to Langflow's new component-constraint system; behavior is unchanged for flow designers.

1.2.x¶

1.2.7¶

New Tool Calls display control — each tool an agent can call has two new knobs: a human-readable title shown in chat and a display mode (hidden, summary line, or expandable line). Set them per tool in the Actions sidebar, or as defaults on a component. Tool Calls →

1.2.6¶

New KARLI passage embeddings — the e5 and q15 models were added to the provider.
New Publishing flows to Karli — a new documentation section walks through it.

1.2.5¶

Improved Chat Input — the component was extended.

1.2.4¶

New dense-experimental-latest — added to the KARLI provider.

1.2.3¶

New KARLI_DATA_EXTRACTION_MOE_LATEST — a new data-extraction model.
Changed Context Input — flows are limited to a single Context Input; the component disables adding more once one is present, with clearer guidance in the docs.
Changed Account menu — the GitHub, Twitter and Discord links were removed.

1.2.2¶

Fixed KARLI model provider — a configuration issue was resolved.

1.2.0¶

Changed Rebased onto Langflow 1.9.0
Improved Documentation site redesign — new layout, navigation and styling, aligned with Karli Studio.
New Split KARLI URLs — the provider supports separate URLs for the LiteLLM gateway and the KARLI API.
New API-docs links — the KARLI and TokenEurope providers link into this documentation.

1.1.x¶

1.1.12¶

New Memory Cell — a new component that reads, writes or deletes a persistent key-value pair scoped to the session, user or agent. Memory Cell →

1.1.11¶

New SHARED_SERVICES_API_KEY — supported by the KARLI provider, alongside improved ReadFile handling.
Fixed KARLI embeddings — the batch-size limit no longer breaks requests.

1.1.10¶

New Build events — components emit build-start and build-end events for observability. (Superseded in 1.2.3.)

1.1.9¶

Fixed KARLI embeddings — a bug caused by unnecessary deserialization was resolved.

1.1.8¶

New Agentlab documentation site — launched, including German translations and screenshots.
New TokenEurope — added as a model provider.
Improved KARLI provider — the LLM lineup was extended.

1.1.7¶

New Identity Input — a new component returning identity information for the user running the flow, decoded from the session token. Identity Input →
New Data Extraction — KARLI data-extraction models and the /extract endpoint, with per-model file-type support (including audio) and JWT-based context for shared services. Data Extraction →

1.1.6¶

Changed Reset-password panel — removed.

1.1.5¶

Changed Rebased onto Langflow 1.8.2
New Context Input — a configurable source selector for agents that perform context-specific lookups. Context Input →

1.1.3¶

New embed-15 — a new embedding model.
Fixed Embedding model selection — picking a model works reliably again.

1.1.0¶

Changed Rebased onto Langflow 1.8.1
New Chat Input chips — chip-based input with reliable routing. Chat Input (Chips) →
New File uploads — the File component accepts user uploads, and files can be sent inline through the API and returned from Chat Input. File Component →
New Provider and bundle gating — available model providers and sidebar bundles can be restricted via the MODEL_PROVIDERS and LANGFLOW_ENABLED_BUNDLES environment variables.
Improved KARLI as an agent provider — with the KARLI icon shown on model components.
Changed Model components — per-component API-key and base-URL inputs are gone; these now come from the central provider settings.
Fixed Input validation — errors return a clean 422 response.

1.0.x¶

1.0.0¶

Changed Rebased onto Langflow 1.7.1
New Fivesquare Docker image — a dedicated image, together with KARLI embeddings configuration.

0.x¶

0.1.0–0.3.0¶

Changed KARLI in the sidebar — added to the sidebar categories and component imports.

0.0.1¶

New First Agentlab release
New KARLI as a model provider — with configurable inputs, API integration and a dedicated icon.
New KARLI Embeddings — including sparse embeddings; the Qdrant component gained an optional vector name and KARLI defaults.
New Publish to Karli — a modal for publishing flows, with user and admin selection.
New Agent flow deletion — plus protected and pre-initialized API keys.
Improved Icon picker — an expanded Material Symbols set and better scrolling behavior.