Imported from adobetoby-maker/tac-brain (AGENTS.md). Install upstream with npx skills add adobetoby-maker/tac-brain. Copyright stays with the author.

AGENTS — Drive's Neural Network Ecosystem

The agents below form an interconnected network sharing OAuth credentials, filesystem memory, and tool access. Each has a defined role and invocation pattern.

Agent 1: Claude Code (TAC)

Role: Primary architect, coder, session orchestrator Identity: ~/.claude/SOUL.md Model: claude-sonnet-4-6 (default) / claude-opus-4-7 (high-stakes) Auth: macOS Keychain "Claude Code-credentials" → sk-ant-oat01-* Bootstrap: /tac skill — syncs memory, shows projects, routes models Strengths: Multi-file reasoning, architecture, tool orchestration, research Memory: ~/.claude/projects/-Users-drive/memory/ (git-backed)

Agent 2: worker-bee-agent / wba (Max OAuth Daemon)

Role: Lightweight autonomous daemon — runs tasks via claude -p (Max OAuth, zero API billing) Identity: ~/.worker-bee/daemon.py Model: claude-sonnet-4-6 (default) | claude-haiku-4-5-20251001 (cron/mechanical) Auth: Max OAuth via Claude Code CLI — same Keychain token, no direct API calls Invoke: wba "task" (inline) | wba -b "task" (background) | wba start (daemon) Strengths: Background tasks, cron, skill context injection, iMessage notify, queue Queue: ~/.worker-bee/tasks/ | Logs: ~/.worker-bee/logs/ | Cron: ~/.worker-bee/cron/jobs.json Key difference from Hermes: uses claude -p as execution engine → Max OAuth works, no exhaustion

Agent 3: Hermes Jr (Max OAuth — Full Hermes Interface)

Role: Full Hermes interface powered by claude -p — profiles, SOUL, cron, gateway, skills Identity: ~/.hermes-jr/SOUL.md | profiles: ~/.hermes-jr/profiles/ Repo: ~/hermes-jr-agent/ → branch hermes-jr (fork of NousResearch/hermes-agent) Model: claude-sonnet-4-6 via Max OAuth (claude -p) — zero API billing Auth: Max OAuth via Claude Code CLI — user:sessions:claude_code scope (works) Invoke: hermes-jr --profile claude -z "task" | hermes-jr --profile sitemanager -z "task" Strengths: Full Hermes CLI interface, profile personas, SOUL injection, memory search Data: ~/.hermes-jr/ | Skills: ~/.claude/skills/ + ~/.hermes/skills/ Key difference from Hermes original: execution engine is claude -p not api.anthropic.com

Side-by-side with Hermes original:

Agent 4: Hermes (Gateway + Complex Long-running)

Role: Persistent autonomous agent, messaging gateway, cron jobs Identity: ~/.hermes/SOUL.md Model: claude-sonnet-4-6 via OAuth (profile: claude) | local LM Studio fallback Auth: Same Claude OAuth (reads from keychain via anthropic_adapter.py) Invoke: hermes --profile claude -z "task" or hermes chat Strengths: Long-running tasks, gateway (iMessage/Telegram/Discord/Slack), cron Memory: ~/.hermes/memory_store.db (SQLite FTS5) + ~/.hermes/team-memory/ Skills: ~/.hermes/skills/ (software-dev, autonomous-ai-agents, etc.)

Active profiles:

Archived local profiles (no LM Studio needed): ~/.hermes/profiles/_archived/ (gemma, granite, qwen, qwenvl)

Agent 3: SiteManager

Role: Autonomous orthobiologic site manager — inventory, orders, monitoring Identity: ~/.hermes/profiles/sitemanager/SOUL.md Model: claude-sonnet-4-6 via Claude OAuth (same keychain as Claude Code + Hermes) Delegation: claude-haiku-4-5-20251001 for mechanical subtasks Sites: orthobiologicpathways.com, ime-coach.com Supplier: LBS Pro Advanced (lbsproadvanced.com) Invoke: hermes --profile sitemanager -z "sync inventory" Cron: Daily inventory sync, 30-min site monitoring Memory: ~/.hermes/profiles/sitemanager/memories/

Credential Flow

macOS Keychain "Claude Code-credentials"
         │
         ├──▶ Claude Code (direct SDK)
         │
         └──▶ Hermes anthropic_adapter.py
                   ├──▶ profile: claude      → claude-sonnet-4-6
                   └──▶ profile: sitemanager → claude-sonnet-4-6
                             (both use same keychain read)

All three agents share ONE Claude OAuth token. No separate API keys. No LM Studio. Token refresh handled by Hermes's credential_pool in ~/.hermes/auth.json.

Memory Sharing

Tool Routing

Dispatch Patterns — How Claude Code and Hermes Work Together

Pattern 1: Explicit Dispatch (you name the agent)

Use when you want a specific background agent to own a task.

# Syntax: dispatch [--bg] [--profile <name>] "<task>"
dispatch --bg "check all worker-bee sites and alert me if anything is down"
dispatch --bg --profile sitemanager "sync lbs inventory to supabase"
dispatch --urgent "send willie elam social post drafts for this week"

The dispatch script is at ~/.hermes/bin/dispatch — queues to ~/.hermes/tasks/pending/.

Pattern 2: Internal Orchestration (let Claude Code decide)

Use when you're mid-session and the task is complex/long. Say naturally: "run this in the background" or "have hermes handle that". Claude Code will:

1. Write the task to ~/.hermes/tasks/pending/<id>.json
2. Spawn hermes --profile claude -z "..." with run_in_background=True
3. Log to ~/.hermes/logs/dispatch-<id>.log
4. Report what was dispatched and the log path

Pattern 3: Cron / Scheduled (autonomous, recurring)

Hermes daemon handles these. Active cron jobs:

Edit cron: ~/.hermes/cron/jobs.json (Hermes reads on daemon start/reload)

Pattern 4: Gateway (iMessage → Hermes → action)

Send a text to your Mac and Hermes acts:

"hermes sync willie elam site"     → Hermes wakes, handles it, texts back
"hermes status"                    → Reports all active agents + last cron runs
"hermes dispatch <task>"           → Same as dispatch CLI but via iMessage

Gateway setup: hermes gateway — runs on bluebubbles/imessage integration.

Willie Elam — Concrete Example

The workflow "update willie elam site and post to social" runs as:

Claude Code (you say it)
  → dispatch --bg "willie elam update: read latest content, draft social posts, iMessage Toby for approval"
  → ~/.hermes/tasks/pending/20260522-173000-abc123.json written
  → hermes --profile claude spawned in background
  → Hermes reads willie-elam site files
  → Drafts LinkedIn + Instagram posts
  → Saves to ~/.hermes/tasks/done/willie-social-20260522.md
  → iMessages Toby: "Here are this week's social drafts. Reply APPROVE to post."
  → You approve via iMessage → Hermes posts (if social API wired)

Adding a New Agent

1. Create ~/.hermes/profiles/<name>/ directory
2. Copy config.yaml from nearest profile (sitemanager for local, claude for cloud)
3. Write ~/.hermes/profiles/<name>/SOUL.md — identity + mission + rules
4. Add to this AGENTS.md with role, model, invoke pattern
5. Add to TAC skill Step 4 quick reference
6. Test: hermes --profile <name> -z "introduce yourself"

Google Cloud / gcloud

• Installed at: /opt/homebrew/bin/gcloud
• Auth status: NOT logged in — run gcloud auth login once when at desktop
• Google OAuth client: [REDACTED-GOOGLE-CLIENT-ID]
• Client secret: [REDACTED-GOOGLE-CLIENT-SECRET]
• Note: Adding redirect URIs to OAuth clients requires Google Cloud Console — no public API
• Automation scripts: ~/.claude/bootstrap/setup-google-oauth.sh, create-supabase-user.sh