Agentic Agents, Testing UI Fixes, and Platform Improvements

April 2, 2026

New Features

Agentic Agents

A new agent type — defined in .agentic.yml files — that runs a multi-step reasoning pipeline autonomously, rather than a single LLM call with tools. Two kinds are supported: Analytics agents answer natural language data questions through an end-to-end pipeline: clarify → specify metrics and dimensions → generate SQL → execute → interpret results.

# analytics.agentic.yml
llm:
  ref: claude-sonnet-4-6
  max_tokens: 16000
databases:
  - my_warehouse
context:
  - ./semantics/*.view.yml
  - ./semantics/*.topic.yml
  - ./workflows/**/*.procedure.yml
states:
  clarifying:
    max_retries: 10
  specifying:
    max_retries: 10
  solving:
    max_retries: 10

App builder agents generate a complete .app.yml Data App from a natural language description.

# app_builder.agentic.yml
llm:
  ref: openai-4o-mini
  max_tokens: 4096
thinking: effort::medium
context:
  - ./semantics/*.view.yml
  - ./workflows/**/*.procedure.yml

Both agent types support all Oxygen connector types (DuckDB, PostgreSQL, BigQuery, Snowflake, ClickHouse) and the Airlayer semantic layer. Analytics agents can also be configured to use an external semantic backend instead of generating raw SQL.

Human-in-the-Loop Suspension

Agentic runs can pause mid-pipeline to ask the user a clarifying question:

Suspension Events — The frontend receives a suspended event when the agent needs additional input before continuing
User Prompt — The UI prompts the user for a response inline within the reasoning trace
Seamless Resume — The run resumes exactly where it left off via POST /analytics/runs/:id/answer, preserving all pipeline state

Agentic Thread UI

The thread detail page now renders agentic runs with a dedicated reasoning interface:

Reasoning Trace — Each pipeline stage is shown in sequence, giving full visibility into how the agent reached its answer
Artifact Sidebar — SQL results, charts, and the final narrative answer are surfaced in a dedicated panel
Suspension Prompt — When the agent needs human input, an inline prompt appears within the thread view
Procedure DAG Panel — When the agent delegates to a workflow, the procedure execution graph is displayed

Agentic reasoning trace showing the clarifying pipeline stage

Agentic analytics result with chart and narrative answer

New API Endpoints

Analytics runs:

Method	Path	Description
`POST`	`/analytics/runs`	Start an analytics run for a given agent and question
`GET`	`/analytics/runs/:id/events`	SSE stream of live reasoning steps and results
`POST`	`/analytics/runs/:id/answer`	Resume a suspended run with a human answer
`GET`	`/analytics/threads/:thread_id/run`	Get run summary with status, answer, and UI event replay

App builder runs:

Method	Path	Description
`POST`	`/analytics/app-runs`	Start an app builder run
`GET`	`/analytics/app-runs/:id/events`	SSE stream of build steps and generated app
`POST`	`/analytics/app-runs/:id/answer`	Resume a suspended build with a human answer
`POST`	`/analytics/app-runs/:id/cancel`	Cancel a running or suspended build

Platform Improvements

Bug Fixes

Testing UI Progress Bar — Fixed a stale state issue where the progress bar did not appear when clicking “Run All” on the Test Dashboard until the page was remounted

​New Features

​Agentic Agents

​Human-in-the-Loop Suspension

​Agentic Thread UI

​New API Endpoints

​Platform Improvements

​Bug Fixes