Edit: github-backup/research/8figureagency-hermes-field-notes-review.md

Note Content (Markdown)

# 8 Figure Agency Hermes Field Notes Review

Source reviewed: `The Dos and Don'ts of Starting with Hermes`, Jordan Ross,
8 Figure Agency, 2026 edition.

## What to Adopt

- Treat every agent like a new hire: define role, purpose, responsibilities,
  non-responsibilities, reporting line, allowed systems and escalation path
  before writing tasks.
- Add role files to the vault design:
  - `soul.md`: identity, role, refusal boundaries and style.
  - `agents.md`: bootstrap file pointing to relevant SOPs.
  - `heartbeat.md`: recurring loop, kept deliberately small.
  - `sops/`: one recurring situation per file.
- Run the analysis phase before building each agent: use a month of real
  artifacts only after permission/privacy is settled, then cluster the 5-7
  recurring situations that represent most of that role's work.
- Use policy, guardrails and hard gates as separate layers. Anything involving
  money, sending, deleting, client-visible output or private data needs a hard
  gate, not just a polite instruction.
- Separate agents by role/lane. Start with one or two named agents, not a
  single all-purpose terminal session.
- Add a later COO/retrospective process that reviews logs and proposes the
  smallest SOP fix instead of rewriting the whole system.

## What to Treat Carefully

- The guide recommends connecting APIs early. For Easier, that should become:
  design first, then read-only access on synthetic or low-risk data, then one
  connector at a time with audit and rollback.
- Telegram can be useful for accessibility, but Anthony prefers Slack. Telegram
  should not be part of the initial plan.
- gbrain sounds aligned with the Markdown-plus-hybrid-search direction, but it
  should be evaluated against QMD/Hermes `llm-wiki` rather than installed
  blindly on the live n8n VM. The current VM has limited disk and memory.
- The "COO agent" should initially produce reports and proposed diffs only. It
  should not autonomously rewrite SOPs until the system has proven itself.

## Revised Easier Sequence

1. Keep the current GCE prep inert.
2. Design one agent out loud, starting with either `Founder EA` or `Research
   Analyst`, because both can be valuable before broad external write access.
3. Add role scaffolding to the vault template: `agents/`, `sops/`,
   `heartbeats/`, and `org-chart.md`.
4. Run on public/synthetic data first.
5. Benchmark memory options: Hermes `llm-wiki`, QMD and gbrain.
6. Add one read-only connector, probably Fathom transcript import or calendar,
   before Gmail/Slack write-capable flows.
7. Only after audit logs and gates exist, consider Slack scheduled jobs.

## Bottom Line

The guide strengthens the recommendation to build an agent organisation, not a
single magical assistant. It does not change the safety conclusion: Easier
should not connect inbox, Slack, payment, CRM or client systems until
the vault, roles, logs and gates are in place.