--- title: Convert all agents to Hermes — feasibility investigation created: 2026-05-26 type: investigation drafted-by: claude-code-coo status: live related: - "[[CLAUDE]]" - "[[findings-2026-05-26]]" - "[[Multi-Hermes-Strategy-2026-05-06]]" --- # CC → Hermes conversion — feasibility > Per Rocky 2026-05-26: "id like to get all the agents converted over to hermes . i think you can do this in hermes, investagate." This is the answer. ## TL;DR **Don't convert everything. Convert SOME. Keep the rest as CC.** The right model is **lane × surface** — each lane (RR, LFCS, COO, SiteOps) gets BOTH a Hermes profile (autonomous, mobile, recurring) AND a CC instance (interactive, code-heavy), with COO orchestrating between them based on task shape. Trying to put everything in Hermes loses CC's strengths (multi-file code editing, plan mode, long-context reasoning, parallel agent fan-out). Trying to put everything in CC loses Hermes's strengths (24/7 cron, Telegram/WhatsApp, persistent memory, autonomous skill execution). ## What each agent does today | Agent | Surface | Primary work shape | Volume | |---|---|---|---| | **claude-code-coo** | CC (laptop) | System + cross-cutting infra: Airtable schema changes, skill catalog evolution, security/pricing checklist iteration, cross-project audits, new skill bootstrapping | High (most-used CC lane) | | **claude-code-opsman** | CC (laptop) | LFCS bid pipeline: per-job runs of `lfcs-project-init`, Phase 0 grill, pricing pack assembly, Drive uploads, AB-Test-Tracker rounds | High (every LFCS bid) | | **claude-code-rr** | CC (laptop) | RateRight CRM/hunt/growth-engine code: hunt script fixes, Supabase migrations, admin UI work | Medium (~weekly cadence) | | **claude-code-orchestrator** | CC (laptop) | Drives `lfcs-pricing-agent` worker pair via append-only file bus. 7-phase state machine. | Medium (per LFCS bid, drives the worker) | | **cowork-lfcs** | Cowork (desktop GUI) | Non-code LFCS work: vault curation, drafting outgoing emails, Airtable record creation, HQ-Vault parallel build for A/B comparison | Medium | | **cowork-hq** | Cowork (desktop GUI) | Cross-cutting HQ work: investigations, audits, curator passes | Medium | | **hermes default** | VPS systemd | LFCS site foreman: daily AM lookahead, EOD summary, weekly programme. Voice transcription. Telegram/WhatsApp interactive. | High autonomous | | **hermes coo** | VPS systemd | (Empty — we're building it) | None yet | | **hermes rr** | VPS systemd | (Empty — never used) | None | | **hermes siteops** | VPS systemd | (Empty — never used) | None | ## What Hermes is genuinely better at 1. **24/7 autonomous cron** — natural-language scheduling, delivery to any platform, skill attachment, job chaining, no-agent script mode with wake gates. CC has nothing equivalent. 2. **Mobile / messaging** — Telegram, WhatsApp, Discord, Slack, SMS (we won't use), Email, Signal, Matrix. Reaches Rocky on his phone in the ute. CC requires terminal. 3. **Persistent memory across sessions** — FTS5 full-text index of every past conversation. By month 2, Hermes knows your decision history better than you do. CC resets every session. 4. **Voice transcription + classification** — Whisper integration, voice notes → structured vault entries. Already wired on `default` profile. 5. **Kanban multi-agent dispatch** — native task board, atomic claim, OS-process-per-task. The hermes-coo plan uses this. 6. **Webhook server** — accept POSTs from external systems (GitHub, Stripe, IoT, cron jobs in other systems). CC has nothing equivalent. 7. **Always-on agent identity** — SOUL.md persists, doesn't get reloaded on every prompt. Lower drift over time. ## What CC is genuinely better at 1. **Multi-file code editing** — Edit/Write tools, plan-mode, parallel tool calls, full bash/PowerShell with rich permission gates. Hermes can shell out but it's clumsy for big refactors. 2. **Long-context reasoning** — 1M context Opus/Sonnet, plus the harness manages context compaction. Hermes has 1M with DeepSeek V4 but the harness around it is less mature. 3. **Interactive design + plan mode** — ExitPlanMode tool, structured planning workflow. Hermes is more autonomous, less interactive. 4. **Subagent fan-out (Agent tool)** — fire 5 parallel investigators returning compressed reports. Today we used this to do the 5-agent vault audit in parallel. Hermes Kanban can do similar but the dev-loop is faster in CC. 5. **Skill ecosystem maturity** — your `lfcs-*` and `lfo-*` skill suites are CC-native, deeply integrated with Airtable MCP + vault file structure + LFCS standards. Porting all to Hermes is months of work for marginal gain. 6. **Voice + image input in chat** — Cowork desktop handles voice/photo/PDF input natively for non-code work. ## Conversion verdict — per lane ### Convert TO Hermes | Current CC role | New Hermes profile | Why | |---|---|---| | Daily ops portion of claude-code-rr (status checks, supabase queries, hunt diagnostics, outreach draft assembly) | `default` (repurposed RR-focused per H-Arch-1) | Skill-shaped, recurring, mobile-relevant. CC was overkill. | | Daily ops portion of claude-code-opsman (RFI status, bid status, Airtable record updates, dockets) | New skills loaded into `default` or a future `hermes-opsman` profile | Recurring, mobile-relevant, doesn't need plan mode | | Cross-cutting strategic triage (the COO role this whole exercise is about) | `coo` (being built now) | Autonomous, cron-driven, scope-cross-project | | Site days work (voice notes, photos, dockets in the van) | `siteops` (activate per H-Arch-1) | Mobile-first by nature | | Recurring monitoring (PM2, scraper rate, vault silence detection, Supabase health) | Hermes crons under `coo` profile | Autonomous by definition | | Voice-brief style AM/EOD pulses | `coo` (strategic) + `default` (operational) | Already what Hermes does best | ### KEEP as CC | CC role | Why not Hermes | |---|---| | **claude-code-coo (system + schema lane)** | Airtable schema changes, skill catalog evolution, cross-project audits. Hermes safety-rail #3 forbids schema changes. Hermes is too autonomous for sensitive infra edits. | | **claude-code-opsman (NEW LFCS bid runs)** | Per-job `lfcs-project-init` is a heavy plan-mode workflow with 11-step pricing pack assembly. Hermes can execute individual skills but plan-mode + parallel-tool-call efficiency makes CC ~5× faster for whole-bid runs. | | **claude-code-orchestrator (drives pricing-agent)** | The append-only file-bus + 7-phase state machine is bespoke CC infrastructure. Hermes Kanban could replace it but rewrite cost is high vs zero benefit (orchestrator works fine today). | | **claude-code-rr (RR code refactors)** | Big code changes (e.g. wire hunt → marketplace, rebuild matchmaker.js, migrate Supabase schema) need plan mode + multi-file edits + 1M context. Hermes can do these but slower + less reliable. | | **cowork-lfcs / cowork-hq** | Non-code work where Cowork's desktop GUI (voice/photo/PDF input, attachment handling, plain English drafting) is its core advantage. Hermes is CLI-shaped. | ## The right architecture — lane × surface ``` ┌─────────────────────────────────────────────────────────────────┐ │ Rocky │ │ (decisions, approvals, strategy direction) │ └────────────────────────┬────────────────────────────────────────┘ │ ┌────────────────────────┴────────────────────────────────────────┐ │ hermes-coo (orchestrator) │ │ reads Inbox, classifies, dispatches, escalates │ │ Telegram pulses (AM/EOD/weekly) via @hq_coo_bot │ └────┬───────────────────┬───────────────────┬────────────────────┘ │ │ │ ┌─────────┴────────┐ ┌────────┴────────┐ ┌────────┴────────┐ │ LFCS lane │ │ RR lane │ │ SiteOps lane │ ├──────────────────┤ ├─────────────────┤ ├─────────────────┤ │ HERMES: │ │ HERMES: │ │ HERMES: │ │ - existing │ │ - default │ │ - siteops │ │ default │ │ profile │ │ profile │ │ crons (AM │ │ (repurposed) │ │ (activated │ │ lookahead, │ │ - DeepSeek V4 │ │ on site │ │ EOD, weekly) │ │ - RR cron │ │ days) │ │ │ │ suite │ │ - voice + │ │ CC: │ │ - mobile/ │ │ photo input │ │ - claude-code- │ │ Telegram │ │ │ │ opsman (per- │ │ │ │ CC: │ │ bid heavy │ │ CC: │ │ - rare; only │ │ workflows) │ │ - claude-code- │ │ when desk │ │ - claude-code- │ │ rr (code │ │ work needed │ │ orchestrator │ │ refactors │ │ on SiteOps │ │ (drives │ │ only) │ │ standards │ │ pricing- │ │ │ │ │ │ agent) │ │ │ │ │ │ │ │ │ │ │ │ COWORK: │ │ COWORK: │ │ │ │ - cowork-lfcs │ │ - cowork-hq │ │ │ │ (drafting, │ │ (audits, │ │ │ │ vault │ │ weekend │ │ │ │ curation) │ │ deep work) │ │ │ └──────────────────┘ └─────────────────┘ └─────────────────┘ │ │ │ ┌─────────┴─────────┐ ┌───────┴────────┐ ┌────────┴────────┐ │ claude-code-coo │ │ │ │ │ │ (cross-cutting │ │ │ │ │ │ system + schema │ │ │ │ │ │ changes, vault │ │ │ │ │ │ structure │ │ │ │ │ │ evolution, │ │ │ │ │ │ Airtable schema) │ │ │ │ │ └───────────────────┘ └────────────────┘ └─────────────────┘ ``` ## Conversion plan (phased, low-risk) ### Phase 1 — Done by tonight - `hermes-coo` SOUL + skills + state pushed to VPS - `default` SOUL swapped to RR-focused per draft - `default` model swapped MiniMax → DeepSeek V4 Pro - LFCS Hermes crons either kept (cheap insurance) or paused (if claude-code-opsman fully covers) ### Phase 2 — Within a week - `siteops` SOUL written + activated on first site day - COO crons live (`coo-eod-pulse` first) - Graphify wired as COO skill, nightly delta-update cron ### Phase 3 — Within a month - Migrate the daily-operations portion of claude-code-rr work into `default` Hermes (skill-shaped tasks) - Migrate the daily-operations portion of claude-code-opsman work into `default` or `hermes-opsman` skills (RFI status, bid pipeline read, status updates) - Tier-1 hibernated asset migration (Radar intel → vault, Cog postmortem → incident playbook) ### Phase 4 — Within a quarter - Evaluate: does CC volume drop because Hermes handles more? If yes, CC becomes "weekend deep work + heavy code only" role - Consider Paperclip if multi-Hermes coordination becomes a UI need (Q3 review) - Consider AION UI on laptop for unified surface ### Phase 5 — Never - Don't try to put `claude-code-orchestrator` into Hermes. The append-only file-bus + state machine is bespoke and working. - Don't put schema changes into Hermes. Lane discipline keeps that in CC. - Don't kill Cowork. It's the GUI tier; complementary. ## What this means for "all agents converted" You won't have fewer "agents" — you'll have: - **Hermes profiles** for autonomous + mobile + recurring work (3 active: `default` RR, `coo`, `siteops` part-time) - **CC instances** for heavy code + interactive design + system maintenance (4 active: `coo`, `opsman`, `rr`, `orchestrator`) - **Cowork sessions** for non-code drafting + voice/photo + vault GUI (2 active: `lfcs`, `hq`) = 9 named agent-roles, but split across 3 surfaces (Hermes/CC/Cowork) and 3 lanes (RR/LFCS/SiteOps). COO orchestrates between Hermes and CC; you coordinate Cowork directly. The net effect: less CC pacing during the day (Hermes handles routine), more Hermes mobile signal in the ute, COO holds the picture, you stop drifting. ## What "converted" success looks like After 30 days: - CC laptop opens 2-3×/day for heavy code work, not 8-10×/day for status checks - Telegram is your daily signal for both RR + LFCS - Hermes-COO catches stalls before they become weeks-long drift - Voice notes from site auto-classify into vault without you opening CC - Worker count is queryable in 2 seconds from your phone That's the model. Convert the SHAPE of the work, not the agent count.