Descriptions:
IndyDevDan demonstrates a “Pi to Pi” two-way agent communication system using multiple GPT 5.5 PI coding agents working as peers rather than in a traditional orchestrator-subordinate hierarchy. The core insight is that agents given a true bidirectional communication channel can collaborate like co-workers, with the best idea winning regardless of which agent produces it.
The video walks through a concrete production engineering scenario: syncing a live database on a Mac Mini to a developer MacBook Pro while stripping personally identifiable information (PII). Two PI coding agents—a production gatekeeper and a dev environment agent—communicate in real time via the Jcoms channel, with the production agent autonomously redacting sensitive data before passing records to its dev counterpart. The workflow mirrors how real engineering teams handle prod-to-dev data migrations.
A second half of the video covers context engineering using E2B and exe.dev sandbox agents, each specialized to understand their respective platform APIs. IndyDevDan argues that focused agents with constrained context windows—he notes a single tool survey hit 10% context (roughly 100K tokens)—outperform monolithic agents given everything. The episode is a practical, code-level look at designing flat multi-agent systems for production use cases where security boundaries and context isolation both matter.
📺 Source: IndyDevDan · Published May 18, 2026
🏷️ Format: Hands On Build
