Building 600+ Connectors w/ AI Airbyte's Automated Connector Factory | Pedro Lopez | PyAI Conf 2026

"What good would an integration platform be if we don't connect to all those different places where your data is at?"

Pedro Lopez, Software Engineer at Airbyte, at PyAI Conf 2026. Previously an early engineer at Grouparoo (acquired by Airbyte) and creator of whatsapp-web.js (20K+ GitHub stars). He breaks down how his team built an automated connector factory — an AI-powered assembly line that builds and maintains 600+ data connectors at scale. Three phases (discovery, build-and-test, acceptance), hard-won lessons on why verification layers matter more than the agent itself, and what happened when they tried handing it all to Claude Code and hoping for the best.

0:00 - Introduction to Airbyte and the connector problem
0:39 - The scale: 600+ replication connectors, 50+ AI agent connectors
1:10 - Can an agent build and maintain all of these?
1:41 - Defining the artifact: declarative YAML connector definitions
2:32 - Why standardizing connector format was a prerequisite for agents
3:18 - The naive approach: hand it to Claude Code and hope for the best
3:57 - Three failure modes: missing knowledge, wrong docs, and fake accountability
4:40 - The assembly line: discovery, build-and-test, and acceptance phases
5:25 - Discovery phase: building context before writing a single line
6:09 - The build-and-test loop: why building without feedback is flying blind
6:47 - Record cassette: testing against live APIs using the VCR pattern
8:29 - Readiness checks: validation the agent cannot game
9:10 - Completeness checks and why prompt-based instructions fell short
9:56 - Trust boundaries: keeping checks outside the agent's control
10:46 - Acceptance phase: does the connector actually solve the user's problem?
11:22 - Golden questions: scenario testing with real user queries
12:01 - Using test traces as PR review artifacts
12:43 - Overview comments and the "front end recording" equivalent for connectors
13:32 - Scaling up: from local Claude Code to Devin to custom background agents
14:22 - Building their own background agent with Ramp Inspect
15:18 - When to let go of control: trust your verification layers first
15:59 - Three takeaways: constrain the problem, add real feedback loops, build ungameable checks
16:34 - Wrap-up

LINKS:
https://airbyte.com/

Видео Building 600+ Connectors w/ AI Airbyte's Automated Connector Factory | Pedro Lopez | PyAI Conf 2026 канала Py AI