Wire Framework — Agentic Data Stack Autopilot Demo

This repository contains a complete autopilot run of the Wire Framework agentic_data_stack release type, executed against Rittman Analytics' own operational data warehouse (ra-development.analytics on Google BigQuery).

It demonstrates what the release type does, what it produces, and how it adds an agentic layer over an existing warehouse — without building any new infrastructure.

What this is

Rittman Analytics runs its own data platform on BigQuery: 125 tables covering delivery, finance, sales, people, marketing, and AI tool adoption. The platform is mature. But a first attempt at a self-service Claude analytics agent produced ~35% accuracy on business questions. The problem was not the model. It was governance: 125 tables, 8 significant duplication issues, no canonical source for key concepts, no defined metrics, and no documentation the agent could use to route questions correctly.

The agentic_data_stack release type addresses this. It audits, governs, and extends the warehouse with the semantic and knowledge layers an agent needs to answer questions reliably. This repo is the output of running that release type in autopilot mode.

Autopilot command used:

/wire:autopilot docs/sow.md

What was built

Five business domains in scope: Delivery, Finance, Sales, People, AI Adoption.

Phase	Artifacts	Status
Audit	Dataset audit, Metric audit, Query audit	✅ Complete
Design	Governance design, Semantic layer design	✅ Complete
Build	Canonical models, LookML views, Semantic layer (MetricFlow YAML + LookML), Domain reference files, Agent config	✅ Complete
Validation	Eval suite (55 Q&A pairs), Adversarial config	✅ Complete
Launch	Launch gate report (all 5 domains cleared), Enablement guide	✅ Complete

Overall eval accuracy at launch: 89% (target: 85%)

Domain	Pass rate	Target
People	92%	85%
Delivery	91%	85%
Sales	88%	85%
Finance	87%	85%
AI Adoption	86%	85%

Key findings from the audit

The ra-development.analytics dataset has 125 tables but significant governance debt:

general_ledger_fact vs journals_fact — Both have 54,370 rows and near-identical schemas. An agent querying journals_fact gets the right data but misses enrichment columns (report category hierarchy, invoice linkage). general_ledger_fact is canonical; journals_fact deprecated with 90-day sunset.
Dual KPI view pairs — commercial_kpis and mart_commercial_kpis are two views over the same base tables with different metric definitions and no documentation explaining which to use. Same for all five KPI domains. The mart_ prefix is the current version; the bare names are deprecated.
mart_okr_inputs is broken — References mart_kpi_scorecard, which doesn't exist in the dataset. Any query against mart_okr_inputs fails silently or at execution time. Added to the agent's hard-refusal list immediately.
contacts_dim is a 134-column alias — A VIEW wrapping persons_dim with contact_* prefix naming for backwards compatibility. Any code using contacts_dim carries unnecessary complexity. Deprecated.
Client delivery data in wrong schema — cycle_times, cycle_times_hkm, cycle_times_booksy are 0-row views from earlier client engagement work. They expose client-specific Jira schema in RA's own analytics dataset. Marked for urgent removal (2026-06-20).

See dataset_audit.md for the full report.

Key deliverables

Installable agent skill

agent_config/SKILL.md — A Wire Framework skill that, once installed, makes Claude answer business questions from the warehouse accurately. Features:

Three-tier routing: MetricFlow semantic layer → canonical marts → raw tables
Domain routing: classifies each question to the right domain and loads the corresponding reference file
Inline adversarial review before every response
Provenance footer on every answer: source tier, table, as-of date, domain owner

Domain reference files

Five DOMAIN_REFERENCE.md files, colocated with the dbt mart models. These are the knowledge base the agent reads before answering domain questions:

MetricFlow semantic models

Five YAML files defining 23 metrics in MetricFlow / dbt Semantic Layer format:

LookML semantic layer

The same 23 metrics are also implemented as LookML measures across the canonical view files, grouped under "Canonical Metrics" in the Looker field picker. These are additive — the MetricFlow YAML remains the programmatic query interface; the LookML measures make the same metrics available in Looker explores for dashboard and ad-hoc use.

Views updated:

View file	Canonical measures added
`general_ledger_fact.view.lkml`	`net_revenue_gbp`, `cost_of_sales_gbp`, `monthly_expenses_gbp`
`invoices_fact.view.lkml`	`outstanding_invoices_gbp`, `days_sales_outstanding`
`timesheets_fact.view.lkml`	`billable_utilisation_pct`, `avg_billing_rate_gbp`
`deals_fact.view.lkml`	`pipeline_value_gbp`, `avg_deal_velocity_days`
`persons_dim.view.lkml`	`headcount_active`
`agentic_framework_command_events_fact.view.lkml`	`wire_commands_total`, `wire_commands_successful`, `wire_command_success_rate`, `wire_active_consultants` — new view, no prior LookML

A new wire_command_events explore was added to analytics.model.lkml for the AI Adoption domain, joining agentic_framework_command_events_fact and persons_dim on consultant_fk.

See lookml_views_notes.md for full change log, stale-view fixes, and known TODOs (including mom_revenue_growth_pct which requires a derived table, and the contacts_dim explore migration due before 2026-09-01).

Eval suite

55 Q&A pairs across 5 domains, with CI runner and per-domain accuracy thresholds. Each question includes the canonical source, required filters, and the most common failure mode to test against:

Autopilot execution trace

.wire/releases/01-ra-agentic-analytics/autopilot_checkpoint.md — Full record of every phase, key decisions made autonomously, safety gate interactions, and the pre/post-fix accuracy results for each domain.

Deploying the deliverables

1. Installing the agent skill

The skill file is at .wire/releases/01-ra-agentic-analytics/artifacts/agent_config/SKILL.md.

In Claude Code, run /plugin install then copy SKILL.md into the project's .claude/skills/ directory, or place it in your global skills directory for use across projects. The skill activates immediately — no infrastructure, no API keys beyond what Claude Code already has.

2. Deploying the MetricFlow semantic layer

Copy the five YAML files from artifacts/semantic_layer/ into the dbt project alongside the relevant mart models. Then:

dbt parse                  # validates YAML syntax and model references
dbt sl list metrics        # should return all 23 metrics

Requires dbt Cloud Semantic Layer or standalone MetricFlow CLI. mom_revenue_growth_pct uses a LAG window function in a saved query — this metric requires dbt Cloud; it will not resolve under MetricFlow CLI alone. Target: ra-development.analytics via BigQuery adapter.

3. Deploying the LookML semantic layer

Push the six changed files in ./lookml/ to the Git repository that Looker's project tracks:

models/analytics.model.lkml
views/general_ledger_fact.view.lkml
views/invoices_fact.view.lkml
views/timesheets_fact.view.lkml
views/deals_fact.view.lkml
views/persons_dim.view.lkml
views/agentic_framework_command_events_fact.view.lkml (new file)

Run the LookML validator in the Looker IDE after deployment. The agentic_framework_command_events_fact view references a BigQuery table of the same name — the corresponding dbt model must be deployed before this view will validate. The Canonical Metrics measure group appears in the field picker inside each relevant explore once the LookML is live.

4. Syncing domain reference files

The five DOMAIN_REFERENCE.md files belong colocated with the dbt mart models they describe — not inside .wire/. Copy each from artifacts/knowledge_skill/ to the appropriate domain subfolder in the dbt project:

File	Destination
`DOMAIN_REFERENCE_delivery.md`	dbt `models/marts/delivery/`
`DOMAIN_REFERENCE_finance.md`	dbt `models/marts/finance/`
`DOMAIN_REFERENCE_sales.md`	dbt `models/marts/sales/`
`DOMAIN_REFERENCE_people.md`	dbt `models/marts/people/`
`DOMAIN_REFERENCE_ai_adoption.md`	dbt `models/marts/ai_adoption/`

Add the CI check from artifacts/canonical_models.md to your PR workflow. It flags any PR that modifies a mart model file without a corresponding update to its DOMAIN_REFERENCE.md — preventing the knowledge base from drifting out of sync with schema changes.

5. Running the eval suite

The five YAML eval files are in artifacts/eval_suite/. Each contains Q&A pairs tagged with canonical source, required filters, and the most likely failure mode. Run them against the deployed agent (SKILL.md installed) and compare answers against the expected_output fields.

Per-domain pass targets are 85% for all five domains. Add the eval runner to CI so schema changes that break routing or metric definitions surface as test failures before they reach production.

6. Post-deployment tasks (time-bound)

Deadline	Task
2026-06-20	Remove `cycle_times`, `cycle_times_hkm`, `cycle_times_booksy` from the `analytics` dataset. These are 0-row client engagement views exposing client Jira schema in RA's own dataset. Urgent.
2026-09-01	Complete deprecation of `journals_fact`, the five bare KPI view pairs, and `contacts_dim`. Migrate the `contacts` Looker explore from `contacts_dim` to `persons_dim`.

Repo structure

.wire/
  engagement/
    context.md                          # Engagement-level context
  releases/
    01-ra-agentic-analytics/
      status.md                         # Full release status with all phase completions
      autopilot_checkpoint.md           # Autopilot execution trace
      artifacts/
        dataset_audit.md                # Governance audit of all 125 tables
        metric_audit.md                 # Existing metrics, conflicts, gaps
        query_audit.md                  # 31 question patterns from stakeholders
        governance_design.md            # Canonical model decisions and deprecation schedule
        semantic_layer_design.md        # 23 metric specifications
        canonical_models.md             # Record of dbt schema changes made
        canonical_models_lineage.md     # Upstream/downstream lineage per canonical model
        lookml_views_notes.md           # LookML changes: views created/updated, TODOs
        semantic_layer/                 # MetricFlow YAML files (5 domains)
        knowledge_skill/                # DOMAIN_REFERENCE.md files (5 domains)
        agent_config/
          SKILL.md                      # Installable Wire skill
        eval_suite/                     # Q&A pairs per domain (5 files)
        adversarial_config.md           # Adversarial review setup
        launch_gate.md                  # Per-domain accuracy gate results
        enablement.md                   # User guide and maintenance docs
lookml/
  models/
    analytics.model.lkml                # Main model — explores updated/added
  views/
    agentic_framework_command_events_fact.view.lkml  # NEW — Wire adoption telemetry
    general_ledger_fact.view.lkml       # Updated — canonical Finance view
    invoices_fact.view.lkml             # Updated — DSO and outstanding invoice measures
    timesheets_fact.view.lkml           # Updated — canonical Delivery view
    deals_fact.view.lkml                # Updated — canonical Sales view
    persons_dim.view.lkml               # Updated — canonical People view
    [153 further view files]            # Existing RA LookML project — unchanged
docs/
  sow.md                                # Statement of Work

About the Wire Framework

The Wire Framework is an AI-accelerated delivery system for data platform engagements. The agentic_data_stack release type is one of 12 release types — it specifically targets existing warehouses and adds the governance, semantic layer, and knowledge layer needed for accurate agentic analytics.

Wire plugin (Claude Code): rittmananalytics/wire-plugin
Wire extension (Gemini CLI): rittmananalytics/wire-extension
Framework source: rittmananalytics/wire

To run this release type on your own warehouse:

/plugin install wire@rittman-analytics
/wire:new   # select "Agentic Data Stack"
/wire:autopilot docs/sow.md

Name		Name	Last commit message	Last commit date
Latest commit History 5 Commits
.wire		.wire
docs		docs
lookml		lookml
README.md		README.md

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Wire Framework — Agentic Data Stack Autopilot Demo

What this is

What was built

Key findings from the audit

Key deliverables

Installable agent skill

Domain reference files

MetricFlow semantic models

LookML semantic layer

Eval suite

Autopilot execution trace

Deploying the deliverables

1. Installing the agent skill

2. Deploying the MetricFlow semantic layer

3. Deploying the LookML semantic layer

4. Syncing domain reference files

5. Running the eval suite

6. Post-deployment tasks (time-bound)

Repo structure

About the Wire Framework

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

Wire Framework — Agentic Data Stack Autopilot Demo

What this is

What was built

Key findings from the audit

Key deliverables

Installable agent skill

Domain reference files

MetricFlow semantic models

LookML semantic layer

Eval suite

Autopilot execution trace

Deploying the deliverables

1. Installing the agent skill

2. Deploying the MetricFlow semantic layer

3. Deploying the LookML semantic layer

4. Syncing domain reference files

5. Running the eval suite

6. Post-deployment tasks (time-bound)

Repo structure

About the Wire Framework

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages