Agent OS — Research Synthesis¶

Source: AI Daily Brief — "How To Build a Personal Agentic Operating System" (Apr 2026) Speaker: Newfar Gaspar (AIDB training programs) Relevance: Directly validates our Hermes persona architecture. We've already implemented most of these layers.

The Seven Layers¶

#	Layer	Purpose	Our Implementation	Gap?
1	Identity	Who you are, rules enforced every session	`~/.hermes/personas/base.md` + 9 role personas	✅ Done
2	Context	What you know (situation-specific knowledge)	`~/context-portfolio/` (10 files)	✅ Done
3	Skills	How you work (reusable instruction sets)	`~/.hermes/skills/` (40+ skills)	✅ Done
4	Memory	What you remember across sessions	3-tier scope system (core/session/ephemeral) + FTS5 + Mnemosyne KG	✅ Done
5	Connections	How agent reaches real systems	Google Workspace, Discord, Spotify, Home Assistant, GitHub	✅ Mostly done
6	Verification	Knowing what to check before shipping	requesting-code-review skill, senior-qa persona, ANDON guard	✅ Partial
7	Automations	Things agents run unattended	20 cron jobs, APScheduler, deterministic runners	✅ Done

8-week shelf life without audit — Personas/context go stale without periodic review. With audit, they compound forever. Our session-retrospective skill handles this but it's not consistently run.
Start read-only on connections — We already do this (Google Calendar = read-only by default, email drafts shown before send). Good.
Automations: only automate manually-validated workflows — We follow this (deterministic-first pattern). Good.
Chief of staff as front door — The video argues this should be the first agent, and it becomes the orchestrator for specialists. Our chief-of-staff persona exists but isn't used as a router/orchestrator. This is a gap.
Dedicated memory for decision logs + relationship context — We have session_search but no structured decision log. Our Mnemosyne KG schema supports it but we don't actively populate decision predicates. This is a gap.
Voice match skill — The video has a dedicated skill for "write like me." We haven't done this. Could be useful for LinkedIn content, email drafts, etc. This is a gap.
Commitment tracker — Track every commitment Adam makes across conversations. We don't do this systematically. This is a gap.
Pre-read skill — One-page briefing before any meeting. We have calendar integration but don't auto-generate meeting prep. This is a gap.

Decision log: Cron or session-end hook that extracts decisions made during conversations → structured format. Queryable later.
Meeting pre-read: Before calendar events, generate a 1-page brief (attendees, recent emails, context from past meetings). Could be a deterministic cron.

Commitment tracker: Skill that scans conversations for "I'll do X by Y" commitments, stores them, and surfaces them in briefings.
Voice match: Curate 5-10 Adam writing samples. Skill references them when drafting emails, LinkedIn, etc. Could integrate into career-coach persona.

Chief of staff as orchestrator: Make chief-of-staff the default persona that routes to specialists. Requires persona-switching at conversation start.
Organisational context files: Stakeholders file, strategy/priorities file. We have the context portfolio structure but haven't populated work-specific context (Jaja org chart, team, etc.)