docs: define relay-native vs subagent collaboration decision

[miyaontherelay]

Summary

• define the shared decision that Relay-native actual agents remain the default model for real multi-agent collaboration
• clarify that harness subagents are execution helpers rather than the core collaboration model
• make the execution-vs-collaboration split explicit for Agent Assistant, Sage, and NightCTO
• capture the proactive GitHub specialist consequence directly

Why

There is an important architectural distinction between:

• helper sub-executions spawned inside a harness, and
• actual collaborating agents with durable role identity, evidence exchange, and proactive/event-driven behavior

This repo needs a clear position so downstream products do not collapse collaboration design into harness convenience.

Key decision

• Relay-native agents are for collaboration
• harness subagents are for execution assistance
• execution backend choice and collaboration topology are separate decisions

Included example

The doc explicitly captures a proactive GitHub-specialist case, where a specialist can receive or be informed about a webhook event like a PR merge, determine user relevance, and produce structured findings or an alert candidate for the user-facing orchestrator.

---

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.