# 11 — HQ Improvements Plan: Lessons from Bhanu's Multi-Agent System

> **Status:** PLAN — Ready for execution
> **Author:** Mario Lavoie
> **Date:** 2026-02-16
> **Source:** [Bhanu's Multi-Clawdbot Video](https://youtu.be/_ISs5FavbJ4) + Antoine's feedback
> **Scope:** Improvements to Atomizer HQ orchestration — not feature dev or project optimization dashboards

---

## Context

Bhanu (founder of SiteGPT) runs 14 OpenClaw agents doing marketing. Key things that work well in his setup and how they map to ours:

| Bhanu's Pattern | Our Equivalent | Status |
|----------------|----------------|--------|
| Single entry point (Jarvis) | Manager orchestrates | ✅ Have it |
| Custom-built task dashboard | — | ❌ **Need this** |
| 15-min polling (agents check for new work) | Heartbeat system | 🔶 Need adaptation |
| Enforced deliverables per task | Handoff JSON schema | 🔶 Need enforcement |
| Squad chat (agents talk freely) | Discord channels | ✅ Have it better |
| Broadcast to all agents | delegate.sh / hooks | ✅ Have it |
| Escalation to Telegram on @mentions | Slack notifications | ✅ Have it |
| Vacation mode (Jarvis auto-approves) | — | 🔶 Nice to have |

**Key insight from Antoine:** We keep Discord as the backbone. Topic-based channels managed by Manager give us *more* visibility than Bhanu's single-dashboard approach. The dashboard supplements Discord — it doesn't replace it.

---

## Improvement 1: HQ Orchestration Dashboard

### What It Is

A lightweight web UI that shows the state of all active work across the Atomizer HQ cluster. Think: a Kanban board + agent status panel. Antoine opens it from any device and sees the full picture.

**This is NOT:**
- The existing Atomizer study/optimization dashboard (that stays separate)
- A feature development tracker (that lives in Discord + roadmap docs)

**This IS:**
- A collaboration/orchestration view for the agent cluster
- Real-time visibility into who's doing what

### Views

#### 1. Task Board (Kanban)

```
┌─────────────┬──────────────┬──────────────┬─────────────┐
│   BACKLOG   │ IN PROGRESS  │   REVIEW     │    DONE     │
├─────────────┼──────────────┼──────────────┼─────────────┤
│ Research    │ 🔧 Tech Lead │ 📋 Secretary │ ✅ Homepage  │
│ CTE data   │ Material     │ Report v2    │   redesign  │
│ [webster]   │ trade study  │ needs CEO    │   [friday]  │
│             │              │ approval     │             │
│ Write test  │ ⚡ Optimizer  │              │ ✅ CTE study │
│ protocol    │ Pareto front │              │   [webster] │
│ [auditor]   │ exploration  │              │             │
└─────────────┴──────────────┴──────────────┴─────────────┘
```

Each card shows:
- Task title + short description
- Assigned agent (emoji + name)
- Time elapsed since assignment
- Status (from handoff JSON: complete/partial/blocked/failed)
- Deliverables attached (if any)
- Blockers (highlighted in red)

#### 2. Agent Status Panel

```
┌──────────────────────────────────────────────┐
│ AGENTS                                        │
│ 🎯 Manager      ● online   3 tasks active    │
│ 📋 Secretary     ● online   1 task active     │
│ 🔧 Tech Lead    ● online   2 tasks active    │
│ 🔍 Auditor      ● online   idle              │
│ ⚡ Optimizer     ● online   1 task active     │
│ 🏗️ Study Builder ○ offline  —                 │
│ 🖥️ NX Expert     ○ offline  —                 │
│ 🌐 Webster      ● online   1 task active     │
└──────────────────────────────────────────────┘
```

Health from `/hooks/health` endpoint per instance. Shows active task count, last activity timestamp.

#### 3. Recent Activity Feed

Chronological log of handoffs, completions, escalations. Like a simplified Discord activity stream but structured:

```
[14:32] 🎯 Manager → 🌐 Webster: "Research CTE of Clearceram-Z HS"
[14:33] 🌐 Webster: accepted task
[14:45] 🌐 Webster: STATUS: complete, CONFIDENCE: high
[14:45] 🎯 Manager → 🔍 Auditor: "Validate Webster's CTE findings"
[14:46] 📋 Secretary → Antoine: "Review needed: Homepage redesign deliverable"
```

#### 4. Escalations & Blockers (Top Bar)

Items waiting for Antoine's input — surfaced prominently so he doesn't need to dig.

### Tech Stack

- **Frontend:** React + Vite + TailwindCSS (same as existing Atomizer dashboard)
- **Backend:** FastAPI endpoint that reads handoff files from `/home/papa/atomizer/handoffs/`
- **Data source:** The handoff JSON from orchestration engine (doc 10) — already structured
- **Refresh:** WebSocket or 10-second polling
- **Auth:** Basic auth or none (local network only)

### Implementation

| Step | Work | Effort |
|------|------|--------|
| 1 | Define task state machine (backlog→active→review→done) | 1h |
| 2 | Add task creation/tracking to orchestrate.sh | 2h |
| 3 | FastAPI endpoints: GET /tasks, GET /agents/status, GET /activity | 3h |
| 4 | React Kanban board component | 4h |
| 5 | Agent status panel + activity feed | 3h |
| 6 | WebSocket live updates | 2h |
| **Total** | | **~15h** |

---

## Improvement 2: Agent Polling for Collaborative Work

### What Bhanu Does

Every 15 minutes, each agent checks the dashboard for new items and adds input if relevant. This creates organic cross-pollination — the SEO agent sees a UX task and adds keyword data without being asked.

### Our Adaptation

**Normal workflow:** Agents poll every **15 minutes** (existing heartbeat cadence). During a poll, agents check:
1. Dashboard task board for tasks in their domain
2. Discord channels they're subscribed to for new context
3. Any handoffs waiting for their input

**Sprint mode:** During active sprints, polling increases to **5 minutes**. Manager activates sprint mode for specific agents working on a deliverable. Sprint mode auto-expires after a configurable duration (default: 2 hours).

### Usage Impact

| Mode | Agents Polled | Frequency | Calls/Hour | Model | Est. Cost/Hour |
|------|--------------|-----------|------------|-------|---------------|
| Normal | 8 | 15 min | 32 | Gemini Flash | ~$0.10 |
| Sprint (3 agents) | 3 | 5 min | 36 | Gemini Flash | ~$0.12 |
| Sprint (all 8) | 8 | 5 min | 96 | Gemini Flash | ~$0.30 |

**Mitigation strategies:**
- Use Gemini Flash for all polling (cheapest model)
- Keep polling context minimal (just task board state, not full history)
- Only sprint-poll *involved* agents, not the whole cluster
- Auto-expire sprint mode to prevent runaway costs

### Implementation

| Step | Work | Effort |
|------|------|--------|
| 1 | Add `sprint-mode.json` config (active agents, frequency, expiry) | 1h |
| 2 | Update agent HEARTBEAT.md templates to check task board | 1h |
| 3 | Manager command: "start sprint [agents] [duration]" | 2h |
| 4 | Dashboard toggle for sprint mode | 1h |
| **Total** | | **~5h** |

---

## Improvement 3: Enforced Deliverables

### What Bhanu Does

Every task requires a deliverable to be marked done. No deliverable = task stays open. This prevents agents from saying "I worked on it" without producing anything.

### Our Adaptation

This is **gold** and we partially have it (handoff JSON has `artifacts` field). We need to make it mandatory and visible.

**Rules:**
1. Every task MUST have a `deliverable_type` defined at creation:
   - `document` — markdown/PDF report
   - `code` — script, config, or code change
   - `analysis` — structured findings with data
   - `recommendation` — decision brief with options
   - `review` — audit/validation report
2. Task cannot move to "done" without an artifact path in the handoff JSON
3. Dashboard shows deliverable status: 📎 attached / ⚠️ missing
4. Manager enforces this — rejects handoffs without deliverables

### Changes to Handoff Schema

```json
{
  "schemaVersion": "1.1",
  "runId": "...",
  "agent": "webster",
  "status": "complete",
  "result": "...",
  "deliverable": {
    "type": "analysis",
    "title": "Clearceram-Z HS CTE Comparison",
    "path": "/home/papa/atomizer/deliverables/2026-02-16-cte-comparison.md",
    "summary": "CCZ HS matches Zerodur Class 0 CTE spec at 0 ± 0.02 ppm/K (20-300K)"
  },
  "confidence": "high",
  "timestamp": "2026-02-16T03:00:00Z"
}
```

### Implementation

| Step | Work | Effort |
|------|------|--------|
| 1 | Update handoff schema (v1.1) with mandatory deliverable block | 1h |
| 2 | Update agent SOUL.md instructions re: deliverables | 1h |
| 3 | Add validation in orchestrate.sh (reject if no deliverable) | 1h |
| 4 | Dashboard: show deliverable status on task cards | 1h |
| **Total** | | **~4h** |

---

## Improvement 4: Discord as the Backbone (Not the Dashboard)

### Why We're Different from Bhanu

Bhanu channels everything through one agent (Jarvis) and uses a custom dashboard as the shared workspace. We have something better: **Discord with topic-based channels**.

**Advantages of our approach:**
- **Visibility** — Antoine sees all agent work in real-time across channels
- **Context** — each channel is a focused workspace, not a noisy firehose
- **Showcase** — looks like a real engineering team (video-worthy, client-facing potential)
- **Direct access** — Antoine can jump into any channel and talk to any agent
- **History** — Discord retains full conversation history per topic
- **Manager still orchestrates** — creates channels, assigns work, enforces deliverables

### What the Dashboard Adds on Top

The dashboard doesn't replace Discord — it's the **bird's eye view**:
- Discord = the workspace (where work happens)
- Dashboard = the control panel (where you see the big picture)

Agents work in Discord channels. The dashboard aggregates their status.

### Manager's Channel Management Role

Manager should actively manage Discord channels:
- Create project channels when new work starts (`#proj-starspec-wfe`)
- Archive channels when work completes
- Pin key deliverables in channels
- Post daily summaries in `#all-atomizer-hq`
- Route new tasks to appropriate channels

### Implementation

| Step | Work | Effort |
|------|------|--------|
| 1 | Update Manager SOUL.md with channel management duties | 1h |
| 2 | Channel naming convention doc | 0.5h |
| 3 | Manager creates channels via Discord API on task creation | 2h |
| **Total** | | **~3.5h** |

---

## Improvement 5: Vacation / Autonomous Mode

### What Bhanu Does

Tells Jarvis "don't wait on me, you be the final call" — Jarvis auto-approves everything for days. Comes back to find 97 items completed.

### Our Adaptation

A "trust level" setting for Manager:

| Level | Behavior |
|-------|----------|
| **Normal** | Escalate decisions, wait for Antoine on approvals |
| **Autonomous** | Manager auto-approves routine work, only escalates high-risk items |
| **Full auto** | Manager approves everything, Antoine reviews async when back |

Activated via: message Manager in Discord or Slack: "Going offline for 3 days, autonomous mode"

### Implementation

| Step | Work | Effort |
|------|------|--------|
| 1 | Add trust-level config to Manager's workspace | 1h |
| 2 | Update Manager SOUL.md with approval logic per level | 1h |
| 3 | Dashboard indicator showing current mode | 0.5h |
| **Total** | | **~2.5h** |

---

## Execution Priority

| # | Improvement | Effort | Impact | Priority |
|---|------------|--------|--------|----------|
| 3 | Enforced Deliverables | 4h | 🔴 High — changes agent behavior now | **Do first** |
| 1 | HQ Orchestration Dashboard | 15h | 🔴 High — visibility into everything | **Do second** |
| 2 | Agent Polling / Sprint Mode | 5h | 🟡 Medium — enables collaboration | **Do third** |
| 4 | Discord Channel Management | 3.5h | 🟡 Medium — better organization | **Do fourth** |
| 5 | Vacation / Autonomous Mode | 2.5h | 🟢 Nice to have | **Do last** |

**Total estimated effort: ~30 hours**

---

## What We're NOT Doing (Intentionally)

- **Single-agent entry point** — We keep Discord multi-channel. More visibility > more control.
- **Custom project management app** — Dashboard is lightweight. Discord is the workspace.
- **Feature development dashboard** — Tracked in Discord + roadmap docs.
- **Project optimization dashboard** — The existing Atomizer dashboard handles study monitoring.
- **Squad chat** — We already have this via Discord channels, and it's better.

---

## Dependencies

- **Orchestration engine (doc 10)** — Handoff JSON schema is the data backbone for the dashboard
- **orchestrate.sh** — Must be working for task tracking
- **Agent cluster** — All 8 instances need to be stable

---

*Prepared by Mario — 2026-02-16*