An experiment from Halo Powered Labs

Halo Workbench icon

Halo Workbench
Agent workflows on your own data.

A desktop workbench where teams of agents answer business questions, assess whether your data is ready, and rebuild dashboards — as workflows you can see, run and inspect. Every figure links to the query that produced it.

Databricks SQL PostgreSQL Windows Runs on your machine
The Halo Workbench window: the Answer a Business Question workflow drawn as a graph of steps, each done, and beside it the run's answer — France finished their chances best — with a table in which every figure carries a link to the query behind it.
One window: the workflow that ran on the left, its answer on the right — each figure tagged with the query it came from. Shown against the World Cup demo database that ships with the app.

A number you can't check is a guess. Halo Workbench shows its work.

The question that matters in the meeting is rarely the one the dashboard anticipated. So it joins an analyst's queue, and what comes back days later is a number in a message thread with no way to check it. A chat assistant is faster, but it answers from memory and guesses at your schema — the figures look right and cannot be verified.

Halo Workbench does the analyst's work as a workflow of specialised agents. One plans the analysis, one writes and runs read-only SQL against your data, one checks the result, one writes the answer. You can watch each step run, and every number in the answer opens the query behind it, explained in business words. When the data doesn't cover the question, the answer says so.

Inspectable

Every figure links to the query that produced it, with the rows at each step and how much of the data it covered.

Composable

Workflows are steps; steps call agents; agents use skills and tools. Build your own agents from the same parts.

Local

One installer. The app, its API and its storage run on your machine — read-only on your data, no service to stand up.

How it works

From a question to an answer you can defend.

Pick a workflow, give it an input, and watch it run. Nothing happens off-screen: each step, each agent and each query is there to open, so a reviewer can follow the work from the words you typed down to the rows that produced the number.

Step 01

Pick a workflow and the data it may read.

A workflow is a job with a defined input: answer a business question, assess a dataset's readiness, turn a dashboard screenshot into a working artifact, or map a saved dashboard onto your columns. Point it at a dataset — the tables you chose from a Databricks SQL warehouse or a PostgreSQL database — and it works over those, nothing wider.

The Workflows catalogue: Answer a Business Question, Assess Data Readiness, Dashboard Image to Artifact and Map a Dashboard to Data, each with its step count, the agents it uses and a Run button.
Four workflows ship with the app; each card names the agents it uses.
Step 02

Watch the steps run — and see what the answer rests on.

The workflow is drawn as a graph and lights up as it runs: receive the question, load the catalog, plan, analyse, validate, write. When it asks you something, it waits for your answer and carries on.

You don't have to watch. Leave the page — or close the window — and the run keeps going; so do a dataset being saved, an eval run or a chart being tailored. The Background tasks list in the sidebar shows what is still working and opens each result when it lands.

The answer ends with a plain coverage line — which tables were read, and whether each query saw the whole table, a ranked top, a sample or a partial result. A number built on a sample cannot pass as a whole-table figure.

The end of a run's answer: a coverage line saying the whole tournaments, matches, match_team_statistics and teams tables were read, the three queries Q1 to Q3 each marked Whole table, the data entities used, and the visuals generated from the answer.
Three queries, each labelled with how much of the data it covered.
Step 03

Click any figure. Read the query behind it in business words.

Every number in an answer carries a tag naming its query. Open it and the query is explained step by step — start from every match, add each team's expected goals, group by team, sort — with the row count at each stage. The SQL and the full result are one tab away, and the whole result downloads as CSV.

It is the query that ran, not a reconstruction. Re-run it against today's data with one click.

The query inspector open over a run: query 2 of 3, marked Whole table, explained in nine numbered steps with row counts, tabs for the result and the SQL, and buttons to download all rows as CSV, re-run and copy the SQL.
The query inspector: nine plain steps from the raw tables to the twelve rows in the answer.
Step 04

Know whether the data deserves your trust before you ask.

Assess Data Readiness profiles a dataset and scores it on the DMBOK data-quality dimensions — completeness, timeliness, consistency, uniqueness and more — then names the weakest dimensions and the most severe findings, table by table and column by column.

Run it once on a new source and you know which answers to lean on and which tables to fix first.

A data readiness run: the six-step workflow graph, and the report — the World Cup dataset scored 93 overall, the Managed band, with timeliness and completeness as the weakest dimensions and a table of the most severe findings.
A readiness score, the weakest dimensions, and the findings behind them.

What's inside

Agents are the building blocks.

Most of the app is not a prompt. It's a library of agents with one job each, and the machinery that keeps their answers honest when the schema is messy and the business vocabulary is not in the column names.

The library

Not one prompt. A team, each with one job.

Olivia plans the analysis. Michael writes and runs the SQL. Sarah assesses data quality. Jessica writes the answer. Hannah explains each query in business words. David repairs a statement the engine rejected; Rachel re-derives the SQL to corroborate a careful answer.

Each agent shows its instructions, skills and tools, and which workflows use it. Build your own from the same parts and try it on your data before you save it.

The Agents library grouped by area — Data Foundations, Analysis, SQL Engineering — with cards for Sarah, Daniel, Kevin, Emily, Michael, Olivia, David, Hannah and Rachel, each with a role and a one-line description.
The agent library, grouped by what each agent is trusted to do.

Dashboards, rebuilt

Drop in a dashboard screenshot and get an interactive artifact back. Then map its widgets and filters onto the columns of your own datasets — and see what is missing.

Your vocabulary, once

Glossary terms, standing instructions, default filters and governed metrics are written once and followed by every answer, ahead of what a model would guess.

Verified queries

A thumbs-up keeps the question and its SQL as a known-good answer that guides the next similar question, so accuracy compounds without a curation project.

Visuals from answers

Turn an answer into an interactive chart built from its own rows, tailor it in plain English, keep every version, and export it with the SQL and data behind it.

Measured, not assumed

Every run is kept with its steps, inputs and output, and eval sets grade answers against known-correct ones — so a change that breaks an answer shows up as a regression.

Bring your own model

OpenAI, Anthropic, LenAI, GitHub Copilot, or any endpoint that speaks the OpenAI or Anthropic API. Keys are stored encrypted on your machine.

Get the app

Download Halo Workbench.

One Windows x64 installer per edition. Download it, run it, and the app sets itself up. Both editions are the same workbench; they differ in name and look, and install side by side.

Halo Workbench

The Halo edition: the Halo Powered look, light and dark.

Windows · .exe installer — x64

Download
Agentic Hub

The Agentic Hub edition: the same workbench in the Agentic Hub look.

Windows · .exe installer — x64

Download

On first launch, open Settings and add your model provider key and a data connection.

Where your data goes

It runs where you run it.

Halo Workbench is a desktop app, not a hosted service. Your warehouse credentials, your runs and your answers stay on the machine that installed it.

Data sources

Databricks SQL warehouses, PostgreSQL databases and REST APIs, queried read-only. A statement that would write is rejected before it reaches the database.

Storage

Datasets, runs, agents, knowledge and settings live in a local database in the app's own data directory.

Leaves the machine

Queries go to the sources you connected; prompts go to the model provider you configured. No telemetry, nothing else.

Credentials

Model provider keys are stored encrypted locally and never rendered back into the interface.

Platforms

Windows x64, in two editions: Halo Workbench and Agentic Hub.

Status

An active Labs experiment — shipping often, and built alongside the teams using it.

Have a question your dashboard can't answer?

That's the one we want. Bring it to the bench and we'll come back with a working answer — and the evidence behind it.

Talk to the Labs