The Workbench

Data Agent Suite

A suite of 10 local-first data agents built around practical data workflows, deterministic evidence, bounded LLM use, and review-ready artifacts.

The Data Agent Suite explores a practical question: where do deterministic checks, review workflows, and LLM support each belong?

The agents start with simple local tools that inspect files and produce traceable outputs. They then build toward richer review workflows where LLMs can help with explanation or suggestions, but not replace the evidence.

What this suite is exploring

Deterministic evidence first

The core facts come from repeatable checks, profiles, comparisons, validations, and trace files.

Bounded LLM support

LLMs are used only where they help with wording, suggestions, or review notes from safe context.

Review-ready artifacts

Each agent writes outputs a person can inspect: reports, traces, findings, summaries, exceptions, or review packs.

Human judgement stays outside the tool

The agents prepare evidence and questions. They do not approve, certify, or replace review.

The journey through the 10 agents

The suite moves from first-pass local checks to fuller review workflows, keeping evidence and judgement separate throughout.

Phase 1

Data foundations

  1. Data Quality Triage Agent

    GitHub repo

    Helps with the first pass of understanding what looks wrong in a dataset.

    Demonstrates: deterministic data quality checks with optional LLM-polished reporting.

  2. Data Reconciliation Agent

    GitHub repo

    Checks whether source-to-target movement preserved the important records and mapped values.

    Demonstrates: mapping-driven comparison, exception outputs, and traceable reconciliation.

  3. Data Dictionary Agent

    GitHub repo

    Creates first-pass data dictionaries from local CSV and XLSX datasets.

    Demonstrates: deterministic profiling as the evidence base, with optional LLM wording support.

Phase 2

Rules, risk, and test design

  1. Data Contract Review Agent

    GitHub repo

    Checks whether a dataset matches agreed contract rules.

    Demonstrates: repeatable validation against required columns, types, nulls, allowed values, and other rules.

  2. Sensitive Field Review Agent

    GitHub repo

    Flags fields that may need careful human review before data is shared, migrated, analysed, or used downstream.

    Demonstrates: policy-led sensitive-field indicators without legal or compliance verdicts.

  3. Data Test Suggestion Agent

    GitHub repo

    Suggests reviewable data tests from profiling evidence and supplied context.

    Demonstrates: LLM-assisted test design with deterministic validation of candidates.

Phase 3

Review workflows

  1. Dataset Onboarding Reviewer Workflow

    GitHub repo

    Prepares an early review pack for a new dataset before people start relying on it.

    Demonstrates: a workflow-shaped agent that moves from profile to context gaps, questions, answers, and report.

  2. Data Quality Investigation Workflow

    GitHub repo

    Helps investigate a stated data quality concern using current and optional baseline files.

    Demonstrates: evidence-led investigation notes without claiming root cause.

  3. Project Evidence Review Agent

    GitHub repo

    Reviews scattered project evidence and highlights what is supported, missing, unclear, or in tension.

    Demonstrates: evidence-pack construction, citation-style review, optional LLM notes, and safe validation.

Phase 4

Migration review

  1. Data Migration Readiness Review Agent

    GitHub repo

    Reviews a migration pack for visible evidence, gaps, mappings, contracts, reconciliation checks, sensitive-field indicators, and test coverage.

    Demonstrates: a fuller review workflow with manifest intake, deterministic artifacts, optional LLM notes, optional LangGraph orchestration, and trace metadata.

Common pattern across the suite

Across the suite, the pattern stays deliberately consistent:

  1. Start with a practical data workflow.
  2. Read local files or supplied evidence.
  3. Run deterministic checks first.
  4. Write traceable artifacts.
  5. Add optional LLM support only after the evidence exists.
  6. Keep human review and final judgement outside the tool.

What changed from agent 1 to agent 10

The early agents focus on one clear job: profile, reconcile, document, or validate.

The later agents become more workflow-shaped. They prepare review packs, manage evidence gaps, structure follow-up questions, and separate deterministic findings from optional LLM notes.

By the final agent, the suite has a fuller pattern: manifest intake, deterministic review stages, aggregated findings, Markdown summaries, optional LLM notes, orchestration seams, and trace metadata.

Where this could go

The next step would not simply be more agents.

The more interesting direction is coordination: a router or workbench layer that can decide which specialist agent fits a given task, pass the right files and context, and keep outputs traceable across workflows.

Before that, each specialist agent needs to remain useful on its own.