Analytics agents built by data teams

A curated list of articles from data teams building analytics agents.

Recommended first reads · Internal builds · Browse by problem · Related reading · Contribute

Recommended first reads

1. Anthropic — How a team makes self-service analytics work. Skills, metric definitions, documentation, and evaluations (Jun 2026), including how the team keeps them current as data models change.
2. OpenAI — Inside an in-house data agent. The architecture behind context preparation, retrieval, permissions, and evaluations (Jan 2026).
3. Ramp — Building one around an existing dbt project. SQL and Jinja examples for context tables and domain documentation (Oct 2025).

Internal builds

Read an article directly from the right column, or click a company for its full reading notes. Dates are publication dates; stacks reflect what the authors used at the time.

Team	Stack	What to read
Alan	Nao, Claude Code, Dust	Adoption across the data team and beyond (Apr 2026)
AngelList	dbt, Snowflake, Devin	Generating context from metadata and code (Aug 2026)
Anthropic	Claude Code, MCP	Skills, documentation, and evaluations (Jun 2026)
Astronomer	DuckDB, Snowflake, Airflow, Slack	Inferring a context layer from metadata and pipeline code (Mar 2026)
Bitrise	PydanticAI, Claude, BigQuery, dbt, Slack	Tools, memory, guardrails, and post-launch lessons (Jul 2026)
BlaBlaCar	dbt, BigQuery, Markdown	Rebuilding a warehouse and its documentation for analytics agents (Sep 2026)
Cloudflare	Trino, R2, DataHub, Workers AI, MCP	A unified data platform and the Skipper analytics agent (May 2026)
Faire	Cursor, Snowflake, Mode	Moving analytics into an IDE (Nov 2025)
GitHub	Copilot, Kusto, Trino	Shared context ownership and PR evaluations (Jun 2026)
Gorgias · 5 posts	dbt, BigQuery, LangSmith, BM25	Cortex platform (Jul 2026) · Context (Mar 2026) · Debugging (May 2026) · Evals (Jul 2026) · Skill retrieval (Aug 2026)
incident.io	Claude Code, Cowork, BigQuery, dbt	Compiling, auditing, and observing an agent’s guidance (Aug 2026)
LinkedIn	LangGraph, LangChain, DataHub	SQL Bot: retrieval, repair, and user experience (Dec 2024)
Meta	SQL, Python	Personal context, iterative analysis, and reusable domain knowledge (Mar 2026)
OpenAI	GPT-5.2, Codex, MCP	Context enrichment and runtime retrieval (Jan 2026)
Ramp · 3 posts	dbt, Snowflake, Looker, Slack	Research overview (Sep 2025) · dbt implementation (Oct 2025) · What changed next (Aug 2026)
Replit	Git	Sharing reviewed corrections between agents (Aug 2026)
Sourcegraph	Gemini, BigQuery, Cloud Functions	DataBot’s tools and feedback loop (Feb 2026)
Uber	GPT-4 Turbo	QueryGPT: domain, table, and column selection (Sep 2024)
Vercel	Claude, AI SDK, Sandbox, Cube	Simplifying the agent around filesystem access (Dec 2025)

Browse by problem

• Giving the agent context: Ramp, Gorgias, AngelList, Astronomer, Bitrise, BlaBlaCar, Cloudflare, incident.io, Meta, OpenAI.
• Testing answers and model changes: Gorgias, Anthropic, ClickHouse, Ramp, LinkedIn, Meta, Uber.
• Fixing mistakes and keeping context current: Bitrise, GitHub, Gorgias, incident.io, Meta, Replit, Astronomer.
• Reducing cost and latency: Cloudflare, Gorgias, Vercel.
• Getting people to use it: Alan, Astronomer, Bitrise, Faire, incident.io, Ramp, LinkedIn, Meta.

Related reading

Product engineering, technical studies, and benchmarks that complement the internal builds.

• Cube: Building an Agentic Analytics Harness · Vendor implementation. Errors, result limits, and permission-aware tools. Stack and notes.

• Omni: Benchmarking Omni’s agentic analytics harness · Vendor implementation. Testing quality, latency, and cost on a vendor’s analytics workload. Stack and notes.

• Cassis: A blank beats a guess · Vendor implementation. Building the first context from existing data assets, with an open-source bootstrap kit. Stack and notes.

• Cassis: Context engineering for analytics agents · Vendor implementation. How we structure tables, metrics, and business rules so an agent can find what it needs. Stack and notes.

• Lyft: From Day 1 to Production · Production engineering. Taking an existing analytics agent from a Streamlit prototype to an authenticated, observable internal application. Stack and notes.

• ClickHouse: The Agentic Analytics Benchmark · Benchmark. A reusable harness built from 201 real questions, with model-jury scoring, cost and latency measurement, and contamination checks. Code · Notes.

• Benchouse: The Analytics Agent Benchmark · Benchmark. Compare accuracy, completeness, restraint, and cost per question across analytics agents. Notes.

Suggest an article

Know a team that has written up its internal build? Open an issue or pull request with the link and a sentence about what makes it useful. See the contribution guide for the format. Corrections and broken-link reports are welcome too.