Manager OS Update: Tool-First Routing, GitHub Change Control & Grey Goo
Updated: 3 hours ago
ARCHIVED ROUTINE UPDATE
This update has been folded into the canonical Manager OS page. It remains published only as historical reference.
Manager OS is getting a proper toolbelt
A useful AI system should not ask another AI to guess something a connected tool can prove directly. The latest Manager OS update formalizes that idea into a tool-first routing layer.
The new default route
1. Retrieve the smallest relevant canonical context. 2. Query the native system that actually owns the fact or artifact. 3. Let the Manager verify and synthesize the evidence. 4. Add a specialist tool or second AI only when it contributes a genuinely different capability or reasoning lens. 5. Verify, then stop when another pass would only duplicate work.
Clear ownership instead of tool soup
Google Drive remains the durable private canon and evidence layer. GitHub now owns engineering change control, technical policies, tests, and rollbackable configuration. The task system owns human actions. Automation tools handle repeatable plumbing. Email, calendar, accounting, and other provider systems are queried directly when they own the live fact.
GitHub graduated from storage to change control
Manager OS now has a private engineering control layer in GitHub with branch-and-review workflows, a machine-readable routing policy, a tool registry, regression checks, and patch tracking. The important boundary is that GitHub does not become a second private-life database. It holds engineering rules and public-safe technical derivatives, while private operational canon stays in its approved source-of-truth system.
Grey Goo now routes by value, not model count
The multi-AI layer is now an escalation path, not the default. A second model is useful for red-teaming, specialist review, or genuinely independent alternatives. Consensus is not evidence, and several models repeating the same unsupported claim still produces one unsupported claim wearing several hats.
What did not change
Manager OS is still designed to reduce cognitive load rather than increase system maintenance. More tools, more automations, and more AI seats are not automatically improvements. New components have to earn a distinct job, respect privacy and authority boundaries, and make the overall system simpler or more reliable.
The practical result: evidence first, tools second, extra brains only when they add something real.
Comments