One file. The whole pipeline.
ADPL is an open JSON-based file format for describing, sharing, and reconstructing agentic data pipelines. A single .adpl file captures the complete pipeline graph, all agent configurations, embedded setup and monitoring prompts, and the human-readable metadata the Pipeline CAD tool needs to reconstruct it on any machine.
Pipelines are hard to share, reproduce, and hand off
Every data team faces the same friction: a pipeline designed in one tool, documented in a wiki, and impossible to reproduce without tribal knowledge. When agents are involved, the problem compounds — no standard way to embed the prompts that set up or monitor the pipeline alongside the pipeline itself.
A single file that is the pipeline
ADPL encodes the full pipeline in one self-contained JSON file. Import it into Pipeline CAD to get the canvas back. Give it to the DE Setup Agent and it knows exactly what to run, check, and report. Share it with a colleague and they have everything — no separate docs, no tribal knowledge.
File structure at a glance
An .adpl file is valid JSON. Six top-level sections tell the complete story of a pipeline — from graph topology to embedded agent prompts.
// pipeline.adpl — a complete agentic pipeline in one file { "adpl": "1.1", // format version "meta": { // human-readable metadata "name": "My Pipeline", "description": "...", "autonomy_level": "L4", // L0 | L2 | L4 "tags": ["kafka", "dbt", "airflow"] }, "pipeline": { // the graph "goal": "Detect anomalies in real-time sensor data", "nodes": [ /* sources · transforms · orchestration · agents · serving */ ], "edges": [ /* directed connections between node output → input ports */ ] }, "agents": { ★ NEW in v1.1 // embedded agent prompts "setup": { /* DE Setup Agent — runs once at deploy */ }, "monitors": [ /* Monitor Agents — run on schedule/trigger */ ] }, "ahi": { // Agent-Human Interface log "log": [ /* observations · alerts · orders · approvals */ ] }, "summary": { // computed snapshot for display "sources": ["kafka"], "transform": "dbt", "orchestration": "dagster", "quality": "great_expectations", "serving": ["dashboard"], "agents": [ /* in-pipeline AI agents */ ] } }
16 node subtypes, 7 categories
Every block on the Pipeline CAD canvas maps to a typed node with a fixed config schema. All config values are strings — making them easy to template, validate, and reference from agent prompts.
postgresql · kafka · mongodb csv · rest_api · minio
dbt · python
airflow · dagster
mcp_server · ai_agent
great_expectations
rest_serve · dashboard · data_warehouse
Embedded agent prompts
Every .adpl file now carries the complete system prompts for the agents that operate around the pipeline — initialising infrastructure, enforcing quality, and monitoring production health. The prompts reference actual config values from pipeline.nodes[].config, so they work without any additional setup.
Four agent types — auto-included based on stack
Three ways to get an .adpl file
⬇ ADPL button. Click it to download the pipeline as a .adpl file, then import it into Pipeline CAD..adpl files with tailored agent prompts referencing the exact config values for each project.Full schema reference and examples
The ADPL specification is published as an open repo. It includes the full README with schema reference tables, the JSON Schema Draft-07 for validation, and ready-to-import example files from real-world pipelines.
.adpl file with the JSON Schema Draft-07. Covers all fields, agent definitions, node subtypes, and port patterns.schema/v1.1.json and use it with ajv, VS Code, or any JSON validator to lint your pipeline files..adpl files with embedded prompts.See ADPL in action
A walkthrough of the Agentic Data Pipeline Language — how a single .adpl file captures an entire pipeline and how Pipeline CAD imports it.