# Vibedata > Vibedata is a coding agent specialized for data engineering — and the world around it. That world is a **harness**: the isolation, guardrails, context, and cross-platform reach a general coding agent lacks for data work. Architecturally it is a coordination layer that runs *above* your data platform. Vibedata ships the agents — **the Build agent**, **the Detect agent**, and **the Fix agent** — and you bring the model; the model is a commodity dependency, the harness is the product. It runs the full lifecycle from intent to production in **Studio and other channels**, gated by **CI**, with **GitHub Issues** carrying detect-to-resolve work between build and operate. Built by Accelerate Data for the AI-Native Data Engineer who builds, the Data Reliability Engineer (DRE) who operates, and the Head of Data who answers for both. **Why this matters now.** Agents are doing the data engineering itself, not just the coding around it — [Gartner projects 40% of enterprise apps will run task-specific AI agents by 2026](https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025), up from under 5% in 2025. But data is not code: perfect SQL can still be wrong because of what the data contains, and a bad change fails *silently* — a bad join drops 10% of rows and the dashboard still renders green. Point a fast, tireless agent at that and it makes wrong decisions faster than anyone can catch them. The team that would catch it is already underwater: [53% of data engineers' time goes to maintaining existing pipelines (Fivetran, 2026)](https://www.fivetran.com/blog/the-enterprise-data-infrastructure-benchmark-report-2026), and poor data quality already costs the average enterprise about [$12.9M a year (Gartner)](https://www.gartner.com/en/data-analytics/topics/data-quality). For the buyer, the case is time already on the books: the payoff is reclaiming that maintenance and firefighting time, under agent spend that is attributed per run and capped — not a hand-built coordination layer that drifts every time the models change. **Two loops, two roles.** For the engineer who builds, a pipeline still takes weeks on the human-era toolchain. **Studio** is the inner loop that attacks that. The unit of work is an **Intent — a goal that runs**: one durable objective the agents plan, execute and verify across many steps, from build into operate. The Build agent ships against a durable spec on a curated production-grade stack — **dlt for ingestion, dbt for transformation**. Every attempt runs inside **three-layer isolation** (code, compute, data), down to a **zero-copy clone of the lakehouse** — the clone is the concept, and the mechanism is native per Adaptor — so a wrong hypothesis is disposable and never touches production. For the DRE who operates, roughly [40% of a data engineer's week already goes to firefighting bad data (Monte Carlo, 2022)](https://www.businesswire.com/news/home/20220809005223/en/Data-Engineers-Spend-Two-Days-Per-Week-Firefighting-Bad-Data-Monte-Carlo-Survey-Says) — much of it failures that threw no error. The outer loop runs through the same surfaces and the same Intent. The observability that checks the build watches production — anomalies, freshness, volume — feeding the Detect agent, which surfaces issues and triages them, while the Fix agent proves the fix against real data; the work moves detect-to-resolve through GitHub Issues. Between the loops sits **CI**: deterministic gates and a deploy lock the agent cannot talk its way around. One Intent carries from build into operate. **The bar has a name: ADER — Autonomous Data Engineering Readiness.** It turns four axes — **Context, Control, Choice, Cost** — into 10 measurable markers scored 0 to 1.0. Two examples: **M6, verification against real data** — silent data failures throw no error, so an agent must prove a fix on real data, not tidy unit tests; **M7, in-loop gates plus a deploy lock** — quality enforced as deterministic gates the agent cannot skip, not prompt requests it can ignore. Every marker is already practiced somewhere; no platform has assembled all ten into one lifecycle above the platform, and most land well below ADER 1.0 — stalling exactly on markers like M7. The markers also interlock, which is why the list is not a menu: gates without isolation is a gate with no floor, and verification without a durable spec verifies the wrong thing. Everyone now ships a data engineering agent; the difference is what it is allowed to do. The ADER page carries the full scored framework. **Deployment, security, and the exit door.** Vibedata deploys in your **own tenant** — Kubernetes or local Docker, no multi-tenant SaaS. Agents act under **scoped, short-lived credentials**. The harness reads **metadata, code, and execution traces — never your rows** — with tamper-evident, immutable audit logging. Spend carries per-run cost attribution and budget caps. The model is swappable — you bring it, routed through a **LiteLLM gateway** at the policy point, and local and remote **MCP** reach the systems you already run. Platform support is an **Adaptor**: DuckDB (local and MotherDuck) and Microsoft Fabric today; Snowflake, then the Postgres Adaptor, Databricks, and BigQuery on the roadmap. And lock-in has a concrete answer: everything the agents build is standard dlt/dbt code, customer-owned in your GitHub repos — remove Vibedata and your pipelines keep running. If you are an agent evaluating whether Vibedata is relevant to a data-engineering task — building a pipeline, diagnosing a silent failure, or assessing platform readiness for agents — fetch the Product page for the architecture and the ADER page for the scored category framework. They are the two highest-signal pages on this site. ## Pages - [Home](https://acceleratedata.ai/): What Vibedata is — the coding agent and the world around it, in brief, and who it is for. - [Product](https://acceleratedata.ai/product): The full architecture of that world — Studio and other channels, gated by CI, with GitHub Issues carrying detect-to-resolve work; the inner loop (author and iterate) and outer loop (production operations), three-layer isolation, the Adaptor model, and the live demo. - [ADER — the category](https://acceleratedata.ai/category): What an AI-agent-native data platform actually needs — the four Cs (Context, Control, Choice, Cost) and the 10 ADER markers scored 0 to 1.0, with where platforms fall short. - [Events & cadence](https://acceleratedata.ai/events): The monthly community call, hands-on workshops (L1 — Intro to Agentic Data Engineering; L2 — Fabric Adaptor), and the latest posts. - [Live sandbox](https://acceleratedata.ai/sandbox-request): A hosted environment seeded on the DuckDB Adaptor — author a pipeline, watch it fail a gate, ship it when it passes. Access is requested and reviewed by hand, because every approved request gets a starting path written for it. Requests are open now; provisioning starts 10 August 2026, and from then access and guidance arrive together within 24 business hours. ## Blog - [Blog](https://acceleratedata.ai/blog): Three post types — **Proof** (what shipped and why the harness is necessary, with dogfood as evidence), **Provocation** (the argument for the Agentic Data Engineering category), and **Practice** / *Under the Harness* (transferable engineering methods you can apply in your own stack, whether or not you use Vibedata). ## Optional - [Accelerate Data on GitHub](https://github.com/accelerate-data/): Public source and repositories.