Skip to content

600-agentic-stack-coordination-v1

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

600-agentic-stack-coordination-v1

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 

README.md

topic agents
type decision
status research-complete
last-validated 2026-05-21
original-query stop agent-as-island pattern and propose coordinated architecture with Claude Code as primary surface (reconstructed)
related-docs 234, 235, 236, 237, 238, 239, 415, 461, 506, 523, 524, 527, 528, 529, 547, 549, 569, 570, 581, 590, 599
tier DEEP

600 — ZAO Agentic Stack Coordination v1

Goal: Take stock of every agentic system Zaal currently runs, identify what works vs what's broken, and propose a coordinated architecture where Claude Code is Zaal's primary surface, Bonfire is shared memory, and specialist bots stay in their lane. Stop the "every agent is its own island" pattern.

Recommendations (no preamble)

Decision Recommendation
Primary surface for Zaal Claude Code CLI. Not ZOE. Not the dashboard. Claude Code is where day-to-day work happens.
Shared memory layer Bonfires (ZABAL bonfire). All agents read from + write to it. Zaal's wallet-gated personal KG.
Bonfire access for agents Direct SDK calls via API key (when Joshua.eth provisions). Until then: Zaal acts as human bridge for non-trivial queries.
ZOE role Concierge for daily tasks + captures + nudges + cross-platform publishing. NOT a code agent. NOT a graph proxy.
Hermes role Code-fix specialist. Triggered by /SHIP FIX or PR webhook. Doesn't need Telegram identity.
ZAOstock Team Bot role Team coordination only. Daily check-ins, leaderboard, group ops. Stays scoped to ZAOstock domain.
Bridge group Passive ingestion surface only. Drop interesting decisions/quotes there → graph captures them. Don't try to do active bot-to-bot orchestration here in v1.
Autonomous bot-to-bot DEFERRED to v2. Wait for Joshua.eth's SDK + MCP server. Don't burn cycles on openclaw tool wiring.

Current Agentic Inventory (verified 2026-05-03)

Tier 1 — Active + reliable

System Surface Stack What it does Health
Claude Code CLI Mac terminal Anthropic Claude (Opus 4.7 1M) Primary dev/research/writing surface for Zaal. This conversation. Solid. Daily driver.
Bonfires bot Telegram DM @zabal_bonfire bonfires.ai Genesis tier Personal KG intake + recall. ~870 nodes, 1200 edges. Confirmed working with 15 traits + system prompt. Green. Verified recall 2026-05-03.
ZAOstock Team Bot Telegram @ZAOstockTeamBot TS + grammy + Supabase, on VPS Onboarding codes, daily check-ins, leaderboard, weekly digest. 17 team members. Green. Live on VPS.
VAULT/BANKER/DEALER agents src/lib/agents in ZAOOS TS + Wagmi/Viem Autonomous trading bots within parameters (autostake, banker, burn, swap). Green. Boring + working.

Tier 2 — Working but underused

System Surface Stack What it does Health
ZOE (@zaoclaw_bot) Telegram DM OpenClaw container, Minimax M2.7 Concierge: daily tasks, captures, ideas, surfaces findings Yellow. Telegram channel was disabled until 2026-05-03 fix. SOUL.md exists but session memory loads slowly.
Hermes coder/critic bot/src/hermes + GitHub PR webhooks Claude CLI subprocess (Max plan auth, no API billing) Autonomous PR fix pipeline. Sprint 1 cost routing live. Green when triggered. Triggers rare (untested loop).
ZAO Devz bot Telegram channel ops bot/src/devz /SHIP FIX command routing to Hermes; channel ops Green.
ZOE learning pings Telegram ZAO Devz topic Python random_tip.py + cron Hourly ZOE-tone tip in ZAO Devz General. Green.

Tier 3 — Paused, broken, or experimental

System Status Notes
Composio AO pilot Paused Doc 415. Installed on BCZ at localhost:3001, 2 PRs open, paused pending Bonfire decision.
Paperclip Live infra, low use paperclip.zaoos.com. Running but not central to current workflow.
OpenClaw ZOEY + WALLET sub-agents Configured, untested Per /vps skill — ZOE dispatches to ZOEY (action) + WALLET (on-chain). No live tasks running through them.
FISHBOWLZ Paused 2026-04-16 Replaced by Juke partnership.
ZOE v2 redesign Brainstorm parked project_zoe_v2_pivot_agent_zero — pivot to Agent Zero proposed but not built.
10-bot fleet plan Brainstorm parked project_tomorrow_first_tasks. Brand-specific bots (Research, Magnetiq, WaveWarZ, POIDH). Not built.
Bot-to-bot bridge group Tried 2026-05-03, not autonomous chat_id -5111907600 wired with requireMention: false for ZOE; Bonfire Group Policy: Open. ZOE doesn't reliably call message.send. Reverting to passive ingestion role.
Hermes Telegram interface Not built Doc 599 wanted this for Pattern 2 (FIX requests via bridge). Deferred.

What We Tried Today + What We Learned (2026-05-03)

1. Bonfires DM ingest pipeline — WORKS

  • 18 YapZ episode .md files uploaded via Document Store → ingested cleanly
  • 8 thematic .md files (q1-wins, zaoos-feature-inventory, people-graph, 5 research-index files) ingested → ~711 facts + 270 edges
  • Verified ChatGPT archive (14 anchors human-triaged) ingested cleanly
  • Recall test passed: bot returns facts WITH source URLs WITH outcome attributes (shipped/dead/evolved) WITH youtube deeplinks

Lesson: Bonfire as PERSONAL KG via DM works. Don't need agent-mediated access for ingest.

2. ZOE telegram channel was disabled — FIXED

  • channels.telegram.enabled was false in openclaw.json since some prior change
  • Patched to true + substituted real bot token from /home/zaal/.env.portal → ZOE responds in Telegram
  • 19+ hours of "Channel is unavailable: telegram" errors stopped

Lesson: Verify config truth. The agent told us "memory reset" earlier when in fact telegram was just disabled — don't trust agent self-reports without VPS verification.

3. Bridge group autonomous bot-to-bot — DEFERRED

  • Created group, added both bots, configured requireMention: false for ZOE in this group, Bonfire Group Policy: Open
  • Token validates, group is real (chat_id -5111907600)
  • BUT: ZOE chose Memory Search (its scout.sqlite local) instead of message.send to bridge to Bonfire. Memory Search then failed.
  • openclaw's tool registry doesn't expose message.send to the LLM cleanly OR ZOE's prompt doesn't route to it

Lesson: Autonomous bot-to-bot is an unsolved tooling problem in openclaw today. Cost to debug > value at this stage. The bridge group still has value as a passive ingestion surface — anything Zaal posts there gets stored.

4. Hermes is alive but rarely triggered — HYGIENE OK

  • Sprint 1 cost routing live (Sonnet/Opus/Haiku)
  • /SHIP FIX in ZAO Devz works
  • Doc 461 4-layer fix-PR pipeline enforces safety
  • But: end-to-end loop (PR breaks → Hermes fixes → PR repaired) hasn't been exercised heavily

Lesson: Hermes is structurally sound. Needs more usage to validate edge cases.

The Architecture Problem

Pattern observed: every agent is an island.

  • ZOE doesn't know what Hermes is doing
  • Hermes doesn't know what ZAOstock bot is doing
  • Bonfire isn't queried by any agent
  • Claude Code (Zaal's daily driver) doesn't have direct Bonfire access either
  • Coordination happens through Zaal's brain, not infrastructure

The pull-toward "make agents talk to each other" is right but premature. We tried Telegram bridge → tooling not ready. SDK not provisioned (waiting on Joshua.eth). MCP server doesn't exist for Bonfire yet.

The usable pattern today: Zaal is the orchestrator. Specialists are tools. Memory is the graph. Bridge group is passive ingest.

Proposed Architecture v1 (works today)

                     ┌──────────────────────┐
                     │  ZAAL (the human)    │
                     │  Orchestrator        │
                     └──────────┬───────────┘
                                │
            ┌───────────────────┼───────────────────┐
            ▼                   ▼                   ▼
   ┌───────────────┐   ┌────────────────┐   ┌──────────────┐
   │ Claude Code   │   │ ZOE @zaoclaw   │   │ Bonfires     │
   │ (terminal)    │   │ (Telegram)     │   │ @zabal_bf    │
   │               │   │                │   │              │
   │ DEV / RESEARCH│   │ DAILY TASKS    │   │ MEMORY       │
   │ WRITING       │   │ CAPTURES       │   │ FACTS        │
   │ HERMES TRIGS  │   │ NUDGES         │   │ DECISIONS    │
   └───────┬───────┘   └────────┬───────┘   └──────┬───────┘
           │                    │                  │
           └────────────────────┼──────────────────┘
                                │
                  ALL THREE WRITE/READ → BONFIRE
                  (via Zaal as proxy until SDK lands)

Lanes locked

Surface Owner Lane
Claude Code CLI Zaal hands-on Code, research, writing, building. Talks to Bonfire via Zaal copy-paste OR via Document Store uploads.
ZOE Telegram Zaal mobile Daily tasks, captures, ideas, content, nudges. Periodically dump captures to Bonfire via Zaal copy-paste.
ZAOstock Team Bot Team Standup, leaderboard, check-ins. Don't touch other domains.
Hermes Triggered by /SHIP FIX or PR webhook Code fixes ONLY. Stays in code lane.
Bonfire DM Zaal direct Knowledge graph queries + ingestion. THE memory layer.
Bridge group Passive surface Anything posted gets ingested. Useful for "I want this remembered" without typing into bot DM.

Communication patterns that ALREADY WORK

  1. Zaal → Bonfire (DM): drag .md file via Document Store → manifest preview → approve. Verified 2026-05-03 with 8 corpora batches.
  2. Zaal → Bonfire (RECALL): DM bot RECALL: <question> → bot returns facts with sources. Verified.
  3. Zaal → ZOE (DM): /add "task", /idea, /capture patterns. Verified ZOE responds.
  4. Zaal → Hermes: /SHIP FIX in ZAO Devz channel. Verified.
  5. Zaal → ZAOstock team bot: team-side commands, /standup etc. Verified.

Communication patterns DEFERRED (don't waste time)

  • ZOE auto-bridging to Bonfire via Telegram (tried today, openclaw tooling not ready)
  • Hermes Telegram identity for bridge participation
  • Cross-bot autonomy (Bot A asks Bot B for info)

These all unlock when Bonfire ships SDK API key + MCP server. Until then, autonomous coordination = Zaal.

Claude Code Integration Strategy

This is the missing piece. Claude Code is Zaal's primary surface but doesn't have Bonfire access.

Today's pattern (manual)

  • Claude Code generates a query: "what do we know about X"
  • Zaal copies to Bonfire DM
  • Bonfire replies
  • Zaal pastes back to Claude Code
  • Claude Code uses

Better pattern (build this week)

Build a small CLI tool: ~/bin/zao-bonfire-recall.sh "<query>". Wraps a curl to Bonfires HTTP API (when key arrives) OR shells out to a Telegram-bot-as-relay we control. Single-line invocation from Claude Code.

~/bin/zao-bonfire-recall.sh "ZABAL contract address with chain"
# → Bonfire DM bot gets the query
# → Reply piped back as text
# → Claude Code uses it inline

This unblocks Claude Code without waiting on full SDK + MCP. Hack today, replace with proper SDK integration once Joshua.eth provisions.

Best pattern (Q3 2026)

Bonfire ships MCP server. Claude Code adds it to .mcp.json:

{
  "mcpServers": {
    "bonfires": {
      "command": "npx",
      "args": ["-y", "@bonfires/mcp-server"],
      "env": { "BONFIRE_API_KEY": "...", "BONFIRE_ID": "..." }
    }
  }
}

Now Claude Code calls mcp__bonfires__recall(query) directly as a tool.

Migration Path (concrete next 4 weeks)

Week 1 — verify today's setup is stable

  • Bonfire DM ingest works (verified 2026-05-03 with 8 corpora batches)
  • ZOE telegram alive (verified 2026-05-03)
  • ZAOstock team bot alive (verified 2026-05-03)
  • Hermes pipeline + cost routing live
  • Bridge group exists as passive surface
  • Doc 590 power-user playbook reviewed by Zaal — daily-use rhythm locked
  • All 3 task #14/#15/#16 (Nexus rebuild / whitepaper rebuild / ZabalSocials rebuild) handed to parallel Claude windows using brief-template pattern

Week 2 — close coordination gaps

  • Build ~/bin/zao-bonfire-relay.sh — Telegram bot relay so Claude Code can RECALL without Zaal copy-paste (interim)
  • Email Joshua.eth with the 8 questions from doc 581 (especially MCP roadmap + API key)
  • Document daily use pattern in CLAUDE.md (/zao-research skill should suggest "RECALL via Bonfire" for any ZAO-context question)

Week 3 — bonfire enrichment

  • Backfill remaining corpora: Farcaster casts (Neynar pull, task #17), publish-history of ZABAL Updates from Paragraph
  • Re-export Bonfire as OWL → quarterly backup at research/agents/bonfire-backups/2026-Q2.rdf

Week 4 — review + iterate

  • Re-evaluate Bonfire vs alternatives (LightRAG self-host per doc 568) at end of 30-day trial
  • If Joshua.eth has provisioned API key by then → build SDK-based agent-to-bonfire integration
  • If not → manual bridge stays, plan for migration

Risks + Mitigations

Risk Mitigation
Bonfire trial expires 2026-05-29 without API key from Joshua.eth Quarterly OWL export protects corpus. Migrate to LightRAG (doc 568) self-host if needed.
ZOE session drift / forgotten SOUL.md sections Quarterly read-and-revise SOUL.md. Add /refresh command if openclaw supports.
Hermes runs wild (auto-PRs that break things) Doc 461 4-layer fix-PR pipeline enforces safety. Not changing.
Multiple agents conflict over Bonfire writes Solo-tenant for now (Q6 grilling locked). Co-founder writes layered later.
Claude Code context bloats with Bonfire data Use RECALL surgically. Don't dump full graphs into context.
New brand bots (Research/Magnetiq/etc) added without coordination Document each new agent in this doc as it ships. Update tier table.

Cost Profile

Component Cost Notes
Anthropic Claude (Max plan) $200/mo Already paid. Powers Claude Code + Hermes.
Bonfires Genesis TBD post-trial Currently free (30-day trial). Pricing pending Joshua.eth.
VPS Hostinger KVM 2 ~$30/mo Already paid. Hosts ZOE + Hermes + ZAOstock bot + Composio AO + Cloudflared.
Telegram $0 All bots use free tier.
Supabase within free tier RLS-backed nexus_links + zaostock-team + others.
Total ~$250-350/mo Below the $500/mo pivot trigger from doc 570

Open Questions for Joshua.eth (from doc 581 + 590)

  1. BONFIRE_API_KEY + BONFIRE_ID + BONFIRE_AGENT_ID — where to get them?
  2. MCP server roadmap — when?
  3. ERC-8004 alignment for agent-authored facts?
  4. OWL/RDF export completeness?
  5. Genesis tier post-trial pricing?
  6. Programmatic graph wipe via SDK?
  7. Idempotency on kengrams.batch re-run?
  8. Dry-run mode for batch ingest?

Also See

  • Doc 547 — multi-agent coordination (predecessor)
  • Doc 549 — Bonfire as second-brain
  • Doc 569 — YapZ ingestion (worked)
  • Doc 570 — agentic memory architecture
  • Doc 581 — bot bug postmortem
  • Doc 590 — power-user playbook
  • Doc 599 — bridge group spec (deferred autonomy)

Next Actions

Action Owner Type By When
Email Joshua.eth with the 8 questions above @Zaal Comms This week
Build ~/bin/zao-bonfire-relay.sh interim Claude-Code-to-Bonfire bridge Claude Code Week 2
Update CLAUDE.md with "use Bonfire RECALL for ZAO context" guidance Claude Doc Week 1
Quarterly OWL backup of Bonfire Claude Process End of June 2026
Re-evaluate Bonfire vs LightRAG at trial end Zaal Decision 2026-05-29
Add new brand bots only after this doc is updated Zaal + Claude Discipline Ongoing

Sources

  • Doc 234-239 (OpenClaw guide, ZOE upgrade history) — internal
  • Doc 461 (fix-PR pipeline) — internal
  • Doc 506 (TRAE AI skip — telemetry concern) — internal
  • Doc 523 (agentic systems audit) — internal
  • Doc 524 (live/archived/started/planned) — internal
  • Doc 527 (multi-bot Telegram coordination playbook) — internal
  • Doc 547 (multi-agent coordination Bonfire-ZOE-Hermes) — internal
  • Doc 549, 569, 570, 581, 590, 599 (Bonfire stack) — internal
  • Lived experience 2026-04-29 → 2026-05-03 (this session) — primary source