The gate decides: work from an agent runs against real data in isolation — tests run, data diffs computed — and the gate's verdict ships two verified changes to production while one returns to the agent with specific feedback for rework
THE AGENTS

Three agents, one accountable lifecycle

Vibedata runs three agents with a chain of accountability, each of which can use scoped subagents and tools — the skills, connectors, sandboxes, and checks bundled as plugins — so one agent can profile a source, generate and test a dbt model, run it in a sandbox, and open a PR without handing off to a different bot. The plugins are extensible: use Vibedata's, pull from the community, or bring your own.

Build

Ships data products — ingestion (dlt) and transformation (dbt) together, from business intent to a deployed, spec-first pipeline, moving through five gated phases (Frame → Design → Build → Verify → Publish). This is the build-side agent the AI-Native Data Engineer directs.

Detect

Watches the platform and reports what it finds — pipeline breakages, quality violations, and improvement opportunities like cost, drift, and missing checks — opening each as an enriched GitHub Issue for the Build agent or the Fix agent to work. It's read-only: it finds problems and raises issues; it never changes your code or your data.

Fix

Answers, investigates, and resolves a question or issue — one loop, three depths: explain, investigate, remediate. Fast and blast-radius-bounded. It never merges its own fix — it asks before it changes anything, and a human merges the PR. Works with the Data Reliability Engineer to keep production healthy.

ARCHITECTURE

How the harness coordinates above the platform

The unit of work is the Intent — a goal that runs. One durable objective the agents plan, execute and verify across many steps, from build into operate; each one carries its own agent, conversation, branch, and isolated workspace. Your data team drives Intents; Vibedata's agent runtime carries each one from goal to production and coordinates the planes that make that safe:

PlaneWhat it does
Ingestion & transformationdlt for ingestion, dbt for transformation — authored, tested, and reviewed inside the loop, as standard code you own and can walk away with.
CI gates & quality enforcementThe checks a production merge cannot skip, plus a deploy lock.
Compute managementWorktree, sandbox, and ephemeral workspace — three-level isolation, down to a zero-copy clone of the lakehouse, so a wrong attempt is disposable. The clone is the concept; the mechanism is native per Adaptor.
Context & knowledgeRuntime context assembled per session from skills, docs, and domain memory that compounds from every incident.
Observability, cost & LLM proxyQuality and spend tracked; a LiteLLM gateway at the policy point for budgets, caps, and model routing. Local and remote MCP reach the systems you already run.

Model choice is a routing decision per lifecycle stage — frontier models where judgment carries the load (specification, design, review), cheaper tiers where the work is well-specified — and the gates hold quality regardless of which tier wrote the change, so you ride the model price curve down.

The whole runtime runs inside your perimeter and reads only metadata, code, and execution traces — never row-level data. Row-level access stays with your agents, inside your environment, on your compute, using your approved model endpoints.

Vibedata architecture: the data team drives Intents; the Vibedata runtime coordinates the three agents and management planes; the domain — your data platform and code — sits below. Vibedata reads only metadata, code, and execution traces, never row-level data.
OPERATING MODES

One lifecycle, many operating modes

The same Intent can be driven different ways — and move between them mid-flight.

Operating modes: how work is triggered — Channel, Automation, or Sense

How work gets triggered

Work is invoked three ways:

  • Channel — a person directs it through a surface (webapp, Slack, Teams).
  • Automation — a schedule or on-demand run.
  • Sense — a signal the platform detects and acts on autonomously.
Operating modes: who's driving — interactive or dispatched, one accountable driver at a time

Who's driving

Every Intent runs in one of two operations:

  • Interactive — a person holds the edit lease and steers live.
  • Dispatched — it runs unattended to completion.

There is one accountable driver at a time, and presence shows who's on — so unattended automation and hands-on work share the same lifecycle without stepping on each other.

LIVE DEMO

Run it before you read any more about it

The sandbox is a hosted, hands-on environment seeded on the DuckDB Adaptor. Run the inner loop — author, iterate, ship — with no install and no tenant setup.

Requests are open now; provisioning starts 10 August 2026. Access comes with a starting path written for you — the scenario to run, the gate to break, and what matters on the platform you use. From 10 August, both arrive within 24 business hours of your request. The recorded walkthrough is open to everyone in the meantime.

Stylized sandbox running the inner loop: author, test against CI gates (all checks pass), and ship via a gated pull request

See it against your own stack

Ready to go deeper? Join the next community call or workshop to talk to the team running it.