Ven Agency — Operating Infrastructure Plan

Version: 1.0
Date: 14 February 2026
Author: Charlie (Chief Operating Systems Architect)
Status: Draft for Andy & Steve review

1. Executive Summary

Ven Agency runs a production AI operations stack that most agencies don't have at all. The Mac Mini + OpenClaw + dual-agent (Charlie/Tom) setup already handles email triage, client reporting, ad management, payroll, and support tickets autonomously. But it has clear single points of failure, manual bottlenecks, and unrealised potential.

This plan upgrades the existing system across five architectural planes — Control, Knowledge, Execution, Interface, and Safety — to reach Minimum Viable Autonomous Agency (MVAA) within 90 days. The core architectural shift is a two-tier agent model: PAs (Charlie, Tom, Sammy, Brian) become pure orchestrators who spawn ephemeral specialist agents per-job. PAs don't do the work — they define it, dispatch it, and relay results. This keeps PAs fast, jobs isolated, and costs controlled.

The key additions:

• CRM/Contact Intelligence Layer — every email, meeting, and interaction feeds a unified contact graph with automatic relationship scoring and meeting prep
• AI Council — monthly multi-persona business reviews that stress-test strategy from 4 angles
• Vector-backed Knowledge Base — ingest URLs, docs, SOPs via Telegram, retrieve with citations
• Meeting → Action Pipeline — transcripts automatically become ClickUp tasks
• API Cost Observatory — real-time spend tracking per agent/workflow with anomaly alerts
• Self-Evolution Engine — daily drift detection against OpenClaw best practices
• Failover Architecture — MacBook Air as warm standby, critical paths replicated

The goal: reduce human involvement in routine operations by ~70%, eliminate bus-factor risk on Andy, and give Steve/James real-time operational intelligence without asking for it.

Estimated implementation cost: ~$200-400/month incremental API spend. No new hardware required (MBA already available as secondary node).

2. Current-State Audit

2.1 System Map

┌─────────────────────────────────────────────────────────────────┐
│                    MAC MINI (PRIMARY HOST)                       │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────────┐   │
│  │ Charlie  │  │   Tom    │  │Sub-Agents│  │  Cron/Launchd │   │
│  │ (Opus)   │←→│ (Opus)   │  │(Codex/   │  │  (15-min      │   │
│  │ Andy ops │  │ Steve ops│  │ Sonnet)  │  │   heartbeats) │   │
│  └────┬─────┘  └────┬─────┘  └────┬─────┘  └──────┬────────┘   │
│       │              │             │               │             │
│  ┌────┴──────────────┴─────────────┴───────────────┴──────┐     │
│  │              OpenClaw Runtime (Node.js)                 │     │
│  │  Skills (50+) │ Memory │ Knowledge Graph │ Sessions     │     │
│  └────────────────────────────┬────────────────────────────┘     │
│                               │                                  │
│  ┌────────────────────────────┴────────────────────────────┐     │
│  │                    Integration Layer                      │    │
│  │  Gmail API │ Calendar │ ClickUp │ Wise │ GA4 │ Meta │    │    │
│  │  GSC │ Google Ads │ GTM │ Cloudflare │ RunCloud │ DO  │    │   │
│  └──────────────────────────────────────────────────────────┘    │
└─────────────────────────────────────────────────────────────────┘
         │              │                │
    ┌────┴────┐   ┌─────┴─────┐   ┌─────┴──────┐
    │Telegram │   │  GitHub   │   │ Cloudflare  │
    │(Control)│   │ (Code)    │   │ Pages (Web) │
    └─────────┘   └───────────┘   └─────────────┘

2.2 Single Points of Failure

2.3 Bottlenecks by Function

Sales (James):

• Lead follow-up is semi-manual — Brian drafts, James reviews via Telegram, but no pipeline visibility
• No automatic relationship scoring or contact enrichment
• Meeting prep is manual (James checks his own notes)

Delivery (Megan + Vietnam team):

• Client report generation is automated but QA is manual
• Ad anomaly detection alerts but doesn't auto-respond
• No unified view of all client health across services

Operations (Andy):

• Email triage automated but approval gates create queues
• Payroll generation automated but verification is manual (rightly so)
• Support tickets flow through but diagnosis is still case-by-case

Business Development (Steve):

• No systematic client health scoring
• No proactive churn risk detection
• Strategic reviews are ad-hoc, not scheduled

3. Future-State Architecture (5 Planes)

3.1 Control Plane — Orchestration, Routing, Cost, Evolution

┌───────────────────────────────────────────────────┐
│                  CONTROL PLANE                     │
│                                                    │
│  ┌─────────────┐  ┌──────────────┐  ┌──────────┐ │
│  │Model Router  │  │Cost Observer │  │Evolution │ │
│  │             │  │              │  │Engine    │ │
│  │ Opus→decisions│ │ Per-agent    │  │Daily     │ │
│  │ Codex→code   │  │ spend track  │  │drift     │ │
│  │ Flash→bulk   │  │ anomaly alert│  │detection │ │
│  │ Sonnet→mid   │  │ weekly report│  │self-patch│ │
│  └──────┬──────┘  └──────┬───────┘  └────┬─────┘ │
│         │                │               │        │
│  ┌──────┴────────────────┴───────────────┴──────┐ │
│  │           Orchestration Bus (OpenClaw)         │ │
│  │  Session mgmt │ Cron scheduler │ Event bus    │ │
│  │  Health checks │ Failover logic │ Audit log   │ │
│  └───────────────────────────────────────────────┘ │
└───────────────────────────────────────────────────┘

Model Router (upgrade existing):

• Current: Manual model selection per task
• Future: Automatic routing based on task classification
  • Classification/extraction/labelling → Gemini Flash ($0.10/1M tokens)
  • Code generation/editing → Codex 5.3 ($0.50/1M tokens)
  • Client-facing writing/strategy → Sonnet 4.5 ($3/1M tokens)
  • Complex decisions/orchestration → Opus 4.6 ($15/1M tokens)
• Implementation: Wrapper function in OpenClaw skills that classifies task complexity before spawning

Cost Observatory (new):

• Intercept layer on all LLM API calls (OpenClaw already logs token counts)
• Script: scripts/cost-observatory.js — parses OpenClaw session logs, aggregates by agent/skill/day
• Cron: Daily summary → memory/costs/YYYY-MM-DD.json
• Cron: Weekly report → Telegram (Steve + Andy)
• Alert: >150% of 7-day rolling average → immediate Telegram notification
• Dashboard: Static HTML on Cloudflare Pages (updated daily)

Evolution Engine (new):

• Daily cron (3am AEDT, low-traffic):
  1. git pull latest OpenClaw docs
  2. Diff against current AGENTS.md, SOUL.md, skill configs
  3. Gemini Flash classifies: breaking change / enhancement / irrelevant
  4. If breaking or high-value enhancement → create ClickUp task + Telegram alert
  5. Log recommendations to memory/evolution/YYYY-MM-DD.md
• Weekly: Self-audit skill usage (which skills unused >30 days? which error frequently?)

3.2 Knowledge Plane — Memory, Vector, Graph, KB

┌───────────────────────────────────────────────────┐
│                 KNOWLEDGE PLANE                    │
│                                                    │
│  ┌──────────┐  ┌──────────┐  ┌─────────────────┐ │
│  │  Memory   │  │ Vector   │  │Knowledge Graph  │ │
│  │  System   │  │ Store    │  │(Entity/Relation)│ │
│  │          │  │          │  │                 │ │
│  │ Daily .md │  │ ChromaDB │  │ graph.json      │ │
│  │ MEMORY.md │  │ or Qdrant│  │ GPT-4o-mini     │ │
│  │ Per-agent │  │ Local    │  │ Incremental     │ │
│  └─────┬────┘  └─────┬────┘  └────────┬────────┘ │
│        │             │                │           │
│  ┌─────┴─────────────┴────────────────┴─────────┐ │
│  │          Unified Retrieval Layer               │ │
│  │  memory_search → vector_search → graph_query  │ │
│  │  Ranked fusion │ Source citations │ Confidence │ │
│  └──────────────────────┬────────────────────────┘ │
│                         │                          │
│  ┌──────────────────────┴────────────────────────┐ │
│  │              KB Ingestion Pipeline             │ │
│  │  Telegram drop → type detect → extract →      │ │
│  │  chunk (512 tokens) → embed → vector store    │ │
│  │  Supported: URL, PDF, DOCX, TXT, MD, HTML     │ │
│  └───────────────────────────────────────────────┘ │
│                                                    │
│  ┌───────────────────────────────────────────────┐ │
│  │           CRM / Contact Intelligence           │ │
│  │  Gmail contacts │ Calendar attendees │ ClickUp │ │
│  │  Dedupe │ Role classification │ Last touch     │ │
│  │  Relationship score │ Meeting prep briefs      │ │
│  └───────────────────────────────────────────────┘ │
└───────────────────────────────────────────────────┘

Vector Store Selection:

• Recommendation: ChromaDB (runs locally on Mac Mini, Python, no external deps)
• Alternative: Qdrant (if scaling beyond Mac Mini later)
• Embedding model: text-embedding-3-small (OpenAI, $0.02/1M tokens — cheapest viable option)
• Storage: ~/.vector-store/ with daily backup to Google Drive

KB Ingestion Pipeline (new skill: kb-ingest):

• Telegram command: /kb <url> or forward any document
• Processing chain:
  1. Type detection (URL → web_fetch, PDF → pdftotext, DOCX → pandoc)
  2. Clean + normalise (strip nav/footer/ads for web)
  3. Chunk at 512 tokens with 50-token overlap
  4. Embed via OpenAI text-embedding-3-small
  5. Store in ChromaDB with metadata (source URL, date, tags)
  6. Confirm in Telegram: "Ingested: [title] — [N] chunks, [source]"
• Retrieval: kb search <query> → top-5 chunks with source citations

CRM/Contact Intelligence (new skill: crm-contacts):

• Daily cron (6am AEDT):
  1. Scan Gmail (last 24h) → extract sender/recipient pairs
  2. Scan Calendar (next 48h) → extract attendees
  3. Dedupe against existing contact store (contacts/contacts.json)
  4. New contacts: Gemini Flash classifies role (client, prospect, vendor, personal, recruiter, spam)
  5. Update "last touch" timestamp + interaction count
  6. Flag contacts with >30 days no touch who are tagged client/prospect
• Meeting prep (triggered 1h before each calendar event):
  1. Look up all attendees in contact store
  2. Pull last 5 email threads with each
  3. Pull any ClickUp tasks mentioning them
  4. Generate 1-paragraph briefing per attendee
  5. Send to Telegram: "Meeting prep for [event] at [time]:"
• Storage: contacts/contacts.json (structured) + vector embeddings for semantic search

3.3 Execution Plane — Agents, Council, Automation

┌───────────────────────────────────────────────────────────────┐
│                     EXECUTION PLANE                            │
│                                                                │
│  ═══════════════════ TIER 1: PAs ═══════════════════          │
│  (Always running, human-facing, orchestration only)            │
│                                                                │
│  ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐            │
│  │ Charlie │ │  Tom    │ │ Sammy   │ │ Brian   │            │
│  │📋 Andy  │ │🔧 Steve │ │📊 Megan │ │🎯 James │            │
│  │ Opus 4.6│ │ Opus 4.6│ │ Sonnet  │ │ Sonnet  │            │
│  │ PA/Orch │ │ PA/Orch │ │ PA/Orch │ │ PA/Orch │            │
│  └────┬────┘ └────┬────┘ └────┬────┘ └────┬────┘            │
│       │           │           │           │                   │
│  ═════╪═══════════╪═══════════╪═══════════╪═══════════════   │
│       │           │           │           │                   │
│  ═══ TIER 2: SPECIALIST AGENTS (Spawned per-job, die after) ══│
│                                                                │
│  Charlie spawns:        Tom spawns:                            │
│  ├─ Onboarder          ├─ Client Health Monitor               │
│  ├─ Payroll Generator  ├─ Support Diagnostician               │
│  ├─ Invoice Processor  └─ Compliance Checker                  │
│  ├─ KB Ingestor                                               │
│  └─ Cost Reporter      Sammy spawns:                          │
│                        ├─ Report Builder (per client)         │
│  Brian spawns:         ├─ Ad Anomaly Investigator             │
│  ├─ Prospect Researcher├─ SEO Auditor                         │
│  ├─ Outreach Drafter   ├─ Creative Analyst                    │
│  ├─ Pipeline Reporter  └─ Competitor Scanner                  │
│  └─ Meeting Prepper                                           │
│                                                                │
│  ┌────────────────────────────────────────────────────────┐   │
│  │         Shared Services (Available to all agents)       │   │
│  │  ChromaDB │ Contact Store │ Cost Observatory │ Gmail    │   │
│  │  Calendar │ ClickUp │ Wise │ Google Ads │ Meta Ads     │   │
│  └────────────────────────────────────────────────────────┘   │
│                                                                │
│  ┌───────────────────────────────────────────────┐           │
│  │              AI COUNCIL (Monthly)              │           │
│  │  Spawned by Tom or Charlie on schedule         │           │
│  │                                                │           │
│  │  ┌──────────┐ ┌──────────┐ ┌──────────┐      │           │
│  │  │ Growth   │ │ Revenue  │ │ Skeptic  │      │           │
│  │  │Strategist│ │ Guardian │ │ Operator │      │           │
│  │  └────┬─────┘ └────┬─────┘ └────┬─────┘      │           │
│  │       │            │            │              │           │
│  │  ┌────┴────┐  ┌────┴────────────┴──────┐      │           │
│  │  │ Team    │  │  Council Moderator     │      │           │
│  │  │Dynamics │  │  (reconcile + rank)    │      │           │
│  │  │Architect│  │                        │      │           │
│  │  └─────────┘  └────────────────────────┘      │           │
│  └───────────────────────────────────────────────┘           │
│                                                                │
│  ┌───────────────────────────────────────────────┐           │
│  │           Cron & Automation Layer              │           │
│  │  Gmail intake (15m) │ Heartbeats (15m)        │           │
│  │  Follow-up tracker │ Maintenance forwarder     │           │
│  │  Cost observatory (daily) │ Contact sync (daily)│          │
│  │  KB ingestion (on-demand) │ Evolution (daily)  │           │
│  │  Council review (monthly) │ Meeting prep (1h)  │           │
│  └───────────────────────────────────────────────┘           │
└───────────────────────────────────────────────────────────────┘

AI Council System (new):

The Council runs monthly (or on-demand via /council). It's a structured multi-perspective business review.

Personas:

Data inputs (auto-collected):

• Revenue data (from Wise transfers + invoicing)
• Client list + service breakdown (from shared client dir)
• Team structure + capacity (from vietnam-team.md + leave-balances.md)
• Ad performance summaries (from Meta/Google APIs)
• Cost observatory data (API spend trends)
• Recent wins/losses (from ClickUp + Gmail)

Flow:

1. Data collection sub-agent gathers inputs → council/data/YYYY-MM.json
2. Each persona runs as a spawned sub-agent (Sonnet 4.5, ~$0.50 each)
3. Outputs collected → fed to Moderator (Opus, ~$2)
4. Final report → council/reports/YYYY-MM.md
5. Summary → Telegram (Steve + Andy) with link to full report
6. Action items → ClickUp tasks (with approval gate)

Estimated cost per council run: ~$4-6

3.4 Interface Plane — Telegram, ClickUp, Gmail, Client-Facing

┌───────────────────────────────────────────────────┐
│                 INTERFACE PLANE                     │
│                                                    │
│  TELEGRAM (Primary Control)                        │
│  ┌─────────────────────────────────────────────┐  │
│  │ Topics:                                      │  │
│  │  #general    — Charlie/Tom direct chat       │  │
│  │  #cron       — Automation status/alerts      │  │
│  │  #council    — AI Council reports            │  │
│  │  #kb         — Knowledge base ingestion      │  │
│  │  #sales      — Pipeline updates, lead alerts │  │
│  │  #support    — Support ticket flow           │  │
│  │  #costs      — API spend reports/alerts      │  │
│  │  #meetings   — Prep briefs, action items     │  │
│  └─────────────────────────────────────────────┘  │
│                                                    │
│  CLICKUP (Task Management)                         │
│  ┌─────────────────────────────────────────────┐  │
│  │ Spaces: Ven Agency │ Clients │ Internal      │  │
│  │ Auto-created: email intake, support tickets, │  │
│  │   council actions, meeting actions, follow-ups│  │
│  │ Statuses: Open → In Progress → Review →      │  │
│  │           Approved → Done                     │  │
│  └─────────────────────────────────────────────┘  │
│                                                    │
│  GMAIL (Communication)                             │
│  ┌─────────────────────────────────────────────┐  │
│  │ Inbound: auto-triage → classify → ClickUp   │  │
│  │ Outbound: draft → Telegram approve → send    │  │
│  │ Labels: For-Charlie, Ann's Clients,          │  │
│  │   Maintenance, Processed, AWAITING REPLY     │  │
│  └─────────────────────────────────────────────┘  │
│                                                    │
│  CLIENT-FACING                                     │
│  ┌─────────────────────────────────────────────┐  │
│  │ Reports: Cloudflare Pages (auto-deployed)    │  │
│  │ Humanizer: ALL external text passes through  │  │
│  │ Approval: No client comms without human sign │  │
│  └─────────────────────────────────────────────┘  │
└───────────────────────────────────────────────────┘

Telegram Topic Structure (upgrade):

• Currently: single chat per agent
• Future: Topic-based group for operational channels
• Each topic has a dedicated purpose, reducing noise in main chat
• Agents post to appropriate topic based on content type
• Implementation: Create Telegram group with topics enabled, update message routing in skills

3.5 Safety & Quality Plane

┌───────────────────────────────────────────────────┐
│              SAFETY & QUALITY PLANE                 │
│                                                    │
│  ┌──────────────┐  ┌──────────────┐              │
│  │ Approval     │  │ Humanizer    │              │
│  │ Gates        │  │ (Mandatory)  │              │
│  │              │  │              │              │
│  │ • External   │  │ All external │              │
│  │   emails     │  │ text passes  │              │
│  │ • Payments   │  │ through AI   │              │
│  │ • Client     │  │ pattern      │              │
│  │   comms      │  │ detection +  │              │
│  │ • Hiring     │  │ rewrite      │              │
│  │ • Public     │  │              │              │
│  │   posts      │  │ Cost: ~$0.01 │              │
│  └──────────────┘  └──────────────┘              │
│                                                    │
│  ┌──────────────┐  ┌──────────────┐              │
│  │ Audit Trail  │  │ Guardrails   │              │
│  │              │  │              │              │
│  │ Every action │  │ • No send    │              │
│  │ logged to    │  │   without    │              │
│  │ session      │  │   approval   │              │
│  │ transcripts  │  │ • No delete  │              │
│  │              │  │   without    │              │
│  │ Searchable   │  │   confirm    │              │
│  │ via memory   │  │ • No payment │              │
│  │ + vector     │  │   >threshold │              │
│  │              │  │ • Rate limits│              │
│  │ Retention:   │  │   on all APIs│              │
│  │ 90 days      │  │ • Sensitive  │              │
│  │ active,      │  │   data never │              │
│  │ archive      │  │   in logs    │              │
│  └──────────────┘  └──────────────┘              │
│                                                    │
│  ┌──────────────────────────────────────────────┐ │
│  │         Verification Layer                     │ │
│  │  • Payroll: mandatory checklist before send   │ │
│  │  • Reports: data accuracy spot-check          │ │
│  │  • Drafts: signature + formatting validation  │ │
│  │  • Sub-agent output: never trust financials   │ │
│  └──────────────────────────────────────────────┘ │
└───────────────────────────────────────────────────┘

Failover Architecture:

PRIMARY: Mac Mini (Ven's Mac mini)
    │
    ├── Heartbeat monitor: Mac Mini → MBA every 5 min
    │   (Script: curl health endpoint via Tailscale)
    │
    ▼
SECONDARY: MacBook Air (Max's host)
    │
    ├── OpenClaw installed, config synced
    ├── Shared skills repo (git pull on failover)
    ├── Google auth tokens replicated
    │
    ▼
FAILOVER TRIGGER:
    3 consecutive missed heartbeats (15 min) →
    MBA auto-starts OpenClaw with failover config →
    Telegram notification: "⚠️ Failover active — MBA running"
    
RECOVERY:
    Mac Mini comes back → health check passes →
    MBA gracefully stops → Telegram: "✅ Primary restored"

Observability Stack:

4. Agentic Org Chart

4.1 Humans

4.2 AI Architecture: Two-Tier Model

Core principle: PAs don't do the work. PAs define and spawn the workers.

Each PA is an agent factory — it knows the specs for every specialist agent it can spawn, routes incoming requests to the right specialist, and reports results back to its human. Specialists are ephemeral: spawned per-job, given clean context, and terminated on completion.

This means:

• PAs stay fast and responsive (never blocked on long tasks)
• Each job gets isolated context (no cross-contamination between clients/tasks)
• Multiple jobs can run in parallel (5 onboardings simultaneously)
• Specialists run on cheap models (Sonnet/Flash), PAs use Opus for routing
• Sarah is retired as a standalone agent — support becomes a specialist agent spawned by Tom

TIER 1: Personal Assistants (Always Running)

Charlie 📋 — Andy's PA

• Purpose: Andy's interface to the entire system. Orchestrates operations, routes work, manages approvals.
• Does NOT: Build reports, research prospects, process invoices, diagnose tickets — spawns specialists for all of these
• What Charlie does directly: Telegram conversations with Andy, quick lookups, approval gate management, agent spawning decisions, memory/context management
• Inputs: Andy's Telegram, Gmail (andy@), Calendar, heartbeat triggers, inter-agent messages, specialist agent results
• Outputs: Telegram responses, approval requests, specialist agent spawns, memory updates
• Specialist library: Onboarder, Payroll Generator, Invoice Processor, KB Ingestor, Cost Reporter, Meeting Prepper, Evolution Auditor
• Guardrails: No external sends without Andy's approval, no payments without confirmation, no deletions without confirmation
• Escalation: If no Andy response in 2h → re-ping; 4h → escalate to Steve via Tom
• Model: Opus 4.6
• Cost tier: High — justified by routing complexity and context management

Tom 🔧 — Steve's PA

• Purpose: Steve's interface. Business ops coordination, client oversight, support orchestration.
• Does NOT: Diagnose support tickets, build client health reports, run compliance checks — spawns specialists
• What Tom does directly: Steve's Telegram, quick decisions, client directory maintenance, inter-agent coordination
• Inputs: Steve's Telegram, Gmail (steve@), Calendar, inter-agent messages from Charlie/Sammy/Brian, specialist results
• Outputs: Telegram responses to Steve, shared directory updates, specialist spawns
• Specialist library: Support Diagnostician, Client Health Monitor, Compliance Checker, SOP Enforcer
• Guardrails: No external sends without Steve's approval
• Escalation: Cross-agent to Charlie for ops issues; to Steve for all approvals
• Model: Opus 4.6
• Cost tier: High

Sammy 📊 — Megan's PA

• Purpose: Megan's interface to marketing data. Routes reporting, monitoring, and analysis requests to specialists.
• Does NOT: Pull API data, build reports, analyse anomalies — spawns specialists
• What Sammy does directly: Megan's Telegram (in plain English, zero jargon), quick campaign questions, routing decisions
• Inputs: Megan's Telegram, scheduled triggers, specialist results
• Outputs: Telegram responses to Megan, specialist spawns, simplified summaries
• Specialist library: Report Builder, Ad Anomaly Investigator, SEO Auditor, Creative Analyst, Competitor Scanner, Copy Extractor, Campaign Builder
• Guardrails: No budget changes without Megan's approval; no ad pausing; alerts only for anomalies
• Escalation: → Megan (ID: 8367326087); → Steve for budget/strategic; → Tom for technical
• Communication style: Plain English for Megan/James, technical for Steve/Tom
• Model: Sonnet 4.5
• Cost tier: Medium
• Workspace: /Users/venagency/sammy/

Brian 🎯 — James's PA

• Purpose: James's interface to sales pipeline. Routes prospecting, outreach, and follow-up work to specialists.
• Does NOT: Research prospects, write emails, build pipeline reports — spawns specialists
• What Brian does directly: James's Telegram, quick pipeline questions, routing decisions, triage
• Inputs: James's Telegram, Gmail/SMS alerts, scheduled triggers, specialist results
• Outputs: Telegram responses to James, specialist spawns, pipeline summaries
• Specialist library: Prospect Researcher, Outreach Drafter, Pipeline Reporter, Meeting Prepper, Follow-up Chaser
• Guardrails: All outbound drafts require James approval; no direct prospect contact
• Escalation: → James (ID: 7679437063); → Steve for strategic prospects
• Communication style: Sharp, energetic, over-prepared
• Model: Sonnet 4.5
• Cost tier: Medium
• Workspace: /Users/venagency/brian/

TIER 2: Specialist Agents (Spawned Per-Job, Ephemeral)

Specialists are spawned by PAs via sessions_spawn. Each gets a focused task, clean context, relevant data, and a timeout. They do the work and return results to the PA, who relays to the human.

Key design rules:

• One specialist per job (not reused across jobs)
• Spawned with only the context they need (no full memory/history)
• Run on the cheapest viable model (Sonnet for analysis, Flash for data collection, Codex for code)
• Hard timeout to prevent runaway costs
• Results always flow back through the PA (specialists never message humans directly)

Charlie's Specialists

Tom's Specialists

Sammy's Specialists

Brian's Specialists

Shared Specialists (Any PA Can Spawn)

AI Council Agents — New (Spawned Monthly)

• Model: Sonnet 4.5 for advisors ($0.50 each), Opus 4.6 for Moderator ($2)
• Cost per run: ~$4-6
• Frequency: Monthly (1st Monday), or on-demand via /council

4.3 Org Chart (Visual)

                 TIER 1: PERSONAL ASSISTANTS
                 (Always on, human-facing)

                    ┌──────────────────┐
                    │   Steve (MD)     │
                    │   + Tom 🔧 PA    │
                    └────────┬─────────┘
                             │
          ┌──────────────────┼──────────────────┐
          │                  │                  │
   ┌──────┴───────┐  ┌──────┴───────┐  ┌───────┴───────┐
   │ Andy (Ops)   │  │ James        │  │ Megan         │
   │ + Charlie 📋 │  │ (Sales)      │  │ (Marketing)   │
   │   PA (Opus)  │  │ + Brian 🎯   │  │ + Sammy 📊    │
   └──────┬───────┘  │   PA (Snnt)  │  │   PA (Snnt)   │
          │          └──────┬───────┘  └───────┬───────┘
          │                 │                  │
  ════════╪═════════════════╪══════════════════╪════════
          │                 │                  │
          │    TIER 2: SPECIALIST AGENTS       │
          │    (Spawned per-job, ephemeral)     │
          │                 │                  │
   ┌──────┴───────┐  ┌─────┴────────┐  ┌──────┴───────┐
   │ Onboarder    │  │ Prospect     │  │ Report       │
   │ Payroll Gen  │  │ Researcher   │  │ Builder      │
   │ Invoice Proc │  │ Outreach     │  │ Ad Anomaly   │
   │ KB Ingestor  │  │ Drafter      │  │ Investigator │
   │ Cost Reporter│  │ Pipeline     │  │ SEO Auditor  │
   │ Meeting Prep │  │ Reporter     │  │ Creative     │
   │ Evolution    │  │ Meeting Prep │  │ Analyst      │
   │ Auditor      │  │ Follow-up    │  │ Competitor   │
   └──────────────┘  │ Chaser       │  │ Scanner      │
                     └──────────────┘  └──────────────┘
   Tom 🔧 spawns:
   ┌──────────────┐
   │ Support      │
   │ Diagnostician│
   │ Client Health│
   │ Monitor      │
   │ Compliance   │
   │ Checker      │
   └──────────────┘

   ┌───────────────────── SHARED ─────────────────────┐
   │ Coder │ Web Researcher │ Doc Drafter │ Data       │
   │ Collector │ ChromaDB │ Contact Store │ Cost Obs   │
   └───────────────────────────────────────────────────┘

   ┌─────────────────────────────────────┐
   │      AI COUNCIL (Monthly/On-Demand) │
   │  Growth │ Revenue │ Skeptic │ Team  │
   │  ──── Council Moderator (Opus) ──── │
   └─────────────────────────────────────┘

5. New Operating Cadence

Daily

Weekly

Monthly

Quarterly

6. Top 10 High-Leverage Automation Blueprints

6.1 Client Onboarding

Trigger: Andy sends /onboard <client name> in Telegram or creates ClickUp task with "Onboarding" tag

Agent chain:

1. Charlie receives command → spawns Onboarder specialist with client details
2. Onboarder creates client folder structure in ~/ven-clients/<client>/
3. Onboarder creates ClickUp onboarding checklist (templated)
4. Onboarder pulls company info (web research, LinkedIn, ABN lookup)
5. Onboarder creates Google Drive folder, shared doc for credentials
6. Onboarder drafts welcome email → returns to Charlie → Charlie presents to Andy for approval
7. Charlie notifies Sammy to spawn tracking setup for new domains

Data touched: ClickUp, ven-clients/, Google Drive, Gmail Approval gates: Welcome email requires Andy approval; ad account access requests require client confirmation Failure modes: Missing client info → Charlie asks Andy; API failures → retry 3x then alert Cost control: Flash for research, Sonnet for drafts (~$0.30 total) Telegram command: /onboard Acme Corp — SEO + Google Ads — contact: jane@acme.com

6.2 Delivery QA

Trigger: Cron (weekly) or manual /qa <client>

Agent chain:

1. Sammy spawns Report Builder specialist for the client
2. Report Builder pulls latest data (Meta, Google, GA4, GSC) via Data Collector
3. Report Builder compares report figures against raw API data
4. Report Builder checks all links, runs mobile rendering test via Smooth CLI
5. Report Builder returns discrepancy report → Sammy creates ClickUp task if needed

Data touched: Cloudflare Pages reports, ad APIs, GA4, GSC Approval gates: None (automated check, human reviews flags) Failure modes: API rate limits → backoff + retry; stale data → flag staleness Cost control: Flash for comparison (~$0.10) Telegram command: /qa all or /qa clientname

6.3 Sales Follow-Up

Trigger: Gmail follow-up tracker (runs on heartbeat, every 15 min)

Agent chain:

1. Brian's heartbeat detects stale AWAITING REPLY tasks
2. Brian spawns Follow-up Chaser specialist per stale lead
3. Chaser checks Gmail for replies in the thread
4. If reply detected → updates ClickUp status → returns to Brian → Brian notifies James
5. If no reply >3 days → Chaser drafts chase email → returns to Brian → Brian presents to James for approval
6. If no reply >7 days → Brian escalates to Steve

Data touched: ClickUp, Gmail, Telegram Approval gates: All chase emails require James approval Failure modes: Gmail API down → retry next heartbeat; thread detection false positive → human review Cost control: Flash for classification, Sonnet for drafts (~$0.05 per run) Telegram command: /followups (show all pending)

6.4 Invoice & Payment

Trigger: Wise webhook or manual /pay <vendor> <amount>

Agent chain:

1. Charlie receives invoice (email attachment or Telegram forward)
2. Extract amount, vendor, currency via Flash
3. Cross-reference with expected payments (vendor list)
4. Create Wise quote → Telegram with details → Andy approves
5. Execute transfer → log to payments/YYYY-MM.json
6. Apply Gmail label, update ClickUp task

Data touched: Wise API, Gmail, ClickUp, payment log Approval gates: ALL payments require Andy confirmation (no auto-pay) Failure modes: OCR extraction error → show raw + ask Andy; Wise API error → retry + alert Cost control: Flash for extraction (~$0.02) Telegram command: /pay LeBlanc 32000000 VND (creates quote for approval)

6.5 Hiring Pipeline

Trigger: New email to careers@ven.com.au or manual /screen <role>

Agent chain:

1. Charlie detects application email
2. Charlie spawns CV Screener specialist with CV attachment + job description
3. CV Screener extracts CV, evaluates against JD criteria
4. CV Screener returns shortlist verdict + reasoning → Charlie creates ClickUp task
5. If strong candidate → Charlie spawns Document Drafter for acknowledgement email → Andy approves
6. If rejected → batch weekly: Charlie spawns Drafter for polite rejections → Andy approves

Data touched: Gmail, ClickUp, recruitment vault files Approval gates: All candidate communications require Andy approval Failure modes: PDF extraction failure → flag for manual review; JD not found → ask Andy Cost control: Flash for initial filter, Sonnet for detailed evaluation (~$0.15 per CV) Telegram command: /screen "Senior Developer" (screen all pending CVs for role)

6.6 Content Pipeline

Trigger: Cron (weekly) or manual /content <client>

Agent chain:

1. SEO audit pulls keyword opportunities (GSC + DataForSEO)
2. Content strategist sub-agent identifies gaps vs competitors
3. Creates content brief with target keywords, headings, word count
4. Draft brief → ClickUp task assigned to Megan
5. Megan reviews/assigns to writer

Data touched: GSC, DataForSEO, competitor sites, ClickUp Approval gates: Content briefs require Megan approval before writing begins Failure modes: API data stale → flag recency; keyword cannibalisation detected → alert Cost control: Flash for data pull, Sonnet for brief (~$0.40) Telegram command: /content "Pure Harvest" — blog ideas for Q2

6.7 Incident Response (Support Tickets)

Trigger: Email to support@ven.com.au (auto-forwarded) or manual /support

Agent chain:

1. Tom detects support email (via heartbeat or forwarding automation)
2. Tom spawns Support Diagnostician specialist with email content + client context
3. Diagnostician classifies severity (P1-P4)
4. Diagnostician creates ClickUp task with formatted description
5. Diagnostician runs initial diagnosis (site health check, error logs via RunCloud)
6. Diagnostician returns proposed fix → Tom presents to Steve for approval
7. On approval, Tom spawns Coder specialist to implement fix → verifies → Tom drafts client response → Steve approves

Data touched: Gmail, ClickUp, RunCloud, client sites, Telegram Approval gates: Fix implementation requires Steve approval; client response requires Steve approval Failure modes: Can't access server → escalate to Andy; diagnosis unclear → escalate with raw data Cost control: Flash for classification, Sonnet for diagnosis (~$0.30) Telegram command: /support status (show all open tickets)

6.8 SOP Enforcement

Trigger: Evolution Engine daily cron (3am AEDT)

Agent chain:

1. Pull latest OpenClaw docs + best practices
2. Diff against workspace files (AGENTS.md, SOUL.md, skills/)
3. Classify changes: breaking / enhancement / irrelevant (Flash)
4. Check skill usage stats (last 30 days)
5. Flag unused skills, outdated configs, missing documentation
6. Create ClickUp tasks for significant updates → Telegram summary

Data touched: OpenClaw docs, workspace files, skill configs, ClickUp Approval gates: Config changes require Andy approval; skill deprecation requires Andy approval Failure modes: Git pull failure → retry; false positive drift → human review Cost control: Flash for classification (~$0.05) Telegram command: /audit (run manual SOP check)

6.9 Client Reporting

Trigger: Cron (monthly, 1st week) or manual /report <client> <period>

Agent chain:

1. Sammy collects data (Meta, Google Ads, GA4, GSC) via ven-data-collector
2. Normalises into report template
3. Generates insights via Sonnet (what changed, why, recommended actions)
4. Humanizer pass on all text
5. Builds interactive web presentation
6. Deploys to Cloudflare Pages
7. Generates email summary → Gmail draft → Andy/Megan approves → sends

Data touched: All ad/analytics APIs, Cloudflare Pages, Gmail, ClickUp Approval gates: Report content requires Megan review; email requires Andy approval Failure modes: API data gaps → flag missing metrics; deploy failure → retry + alert Cost control: Flash for data collection, Sonnet for insights, ~$1-2 per report Telegram command: /report "MRH" feb-2026

6.10 Ad Anomaly Response

Trigger: Anomaly detection cron (runs 3x daily)

Agent chain:

1. Sammy pulls performance data (Meta + Google)
2. Compares against rolling 7-day averages
3. Detects: spend spikes, CPA increases >25%, CTR drops >15%, budget exhaustion, creative fatigue
4. If anomaly detected: a. Root cause analysis (audience saturation? creative fatigue? competitor pressure? tracking issue?) b. Recommended action (pause creative, adjust budget, swap audience, check tracking) c. Alert → Telegram with severity + recommendation
5. If P1 (>50% spend anomaly or tracking failure) → immediate alert to Andy + Megan
6. Creates ClickUp task with full analysis

Data touched: Meta Marketing API, Google Ads API, GA4 (for conversion verification) Approval gates: No automated campaign changes — all recommendations require Megan approval Failure modes: API timeout → retry; false positive → confidence threshold (>80% to alert) Cost control: Flash for data pull + comparison, Sonnet for root cause (~$0.20) Telegram command: /anomaly check (run manual scan)

7. SOP & Knowledge Refactor Plan

7.1 Where SOPs Live

Rule: If it's executable by an AI agent → skill. If it's reference for humans → ClickUp Doc. If it's both → skill with ClickUp link.

7.2 KB Ingestion Pipeline

INPUT                    PROCESS                   STORAGE
─────                    ───────                   ───────
Telegram /kb <url>  ──→  web_fetch + clean    ──→  ChromaDB (vectors)
Telegram forward doc ──→  pdftotext/pandoc    ──→  + metadata JSON
Gmail attachment    ──→  type detect          ──→  + source citation
Manual file drop    ──→  chunk (512 tok)      ──→  Backed up to
                    ──→  embed (text-embed-3) ──→  Google Drive (daily)
                    ──→  store + index        ──→

RETRIEVAL
─────────
Query ──→ 1. memory_search (MEMORY.md + memory/*.md)
      ──→ 2. vector_search (ChromaDB, top-5)
      ──→ 3. graph_query (knowledge graph, entity relationships)
      ──→ Ranked fusion (dedup + confidence score)
      ──→ Response with [Source: path#line] citations

Skill: kb-ingest (new)

• Commands: /kb <url>, /kb search <query>, /kb stats, /kb list
• Storage: ~/.vector-store/chroma/
• Embedding cost: ~$0.01 per 50 pages
• Daily backup: rsync to Google Drive

7.3 Retrieval Logic

Three-tier retrieval (executed in parallel, results fused):

1. Memory search (existing) — semantic search over MEMORY.md + daily files

   • Best for: recent context, decisions, personal preferences
   • Speed: <1s

2. Vector search (new) — ChromaDB similarity search

   • Best for: ingested documents, SOPs, reference material
   • Speed: <2s
   • Returns: top-5 chunks with source + confidence

3. Knowledge graph (existing) — entity/relationship traversal

   • Best for: "who is related to what", cross-referencing people/projects
   • Speed: <3s

Fusion: Results ranked by relevance score, deduplicated, top-10 presented with citations.

7.4 Drift Detection

Evolution Engine (daily cron, 3am AEDT):

1. OpenClaw best practices:

   • git pull latest docs from OpenClaw repo
   • Compare prompting patterns, config recommendations, new features
   • Flag applicable improvements

2. Workspace health:

   • AGENTS.md last modified >30 days? → reminder
   • Skills with >5 consecutive errors? → flag for review
   • Skills unused >60 days? → suggest deprecation
   • Memory files >30 days old in HOT tier? → suggest archival

3. Config drift:

   • Compare running OpenClaw config against documented expected state
   • Flag any undocumented changes

4. Output: memory/evolution/YYYY-MM-DD.md + Telegram alert if action needed

7.5 Workspace File Evolution

8. 90-Day Implementation Roadmap

Phase 1: Foundation (Days 0–30)

Goal: Cost visibility, failover readiness, KB infrastructure

Phase 1 milestone: Cost tracking live, KB accepting URLs, meeting prep running, failover tested.

Phase 2: Intelligence (Days 31–60)

Goal: AI Council, Sales Intelligence, enhanced KB

Phase 2 milestone: AI Council run completed, Sales Intel tracking pipeline, Sammy monitoring ads, KB handling multiple doc types.

Phase 3: Autonomy (Days 61–90)

Goal: Self-evolution, full cadence, MVAA

MVAA Milestone (Day 90):

• ✅ All routine email handled automatically (human approves, doesn't draft)
• ✅ Client reports generated end-to-end (human reviews, doesn't build)
• ✅ Ad anomalies detected and recommended actions surfaced (human decides, doesn't investigate)
• ✅ Sales pipeline tracked and follow-ups automated (human closes, doesn't chase)
• ✅ Support tickets triaged, diagnosed, and fix-proposed (human approves, doesn't diagnose)
• ✅ Monthly AI Council producing actionable business insights
• ✅ Cost tracking and model routing optimised
• ✅ Failover tested and ready
• ✅ KB accepting and retrieving knowledge on demand
• ✅ Self-evolution flagging drift and suggesting improvements

9. Risk Register & Controls

10. Minimum Inputs Still Needed From Ven

From Andy

From Steve

From James

Assumptions Made (Flag for Review)

1. MBA is available as failover — assumed from TOOLS.md reference to Max's MBA host
2. ChromaDB can run on Mac Mini — confirmed compatible (Python, macOS arm64)
3. Monthly API budget ~$200-400 — based on current spend patterns + new capabilities
4. Megan is open to AI-assisted monitoring — not confirmed, assumed from existing anomaly-detection skill usage
5. James uses ClickUp for pipeline — not confirmed, may use spreadsheet or mental model
6. Meeting transcripts available — assumed Fathom or equivalent is in use or planned
7. All Google APIs remain on current auth — no new OAuth scopes needed for contact sync (already have Gmail + Calendar)
8. Steve reads Telegram regularly — council reports and support approvals route there
9. Vietnam team doesn't need direct AI agent access — work assignments come through Andy/Steve
10. Client fees are trackable — via Wise transfers or invoice records

Appendix A: Cost Estimates

Appendix B: File/Directory Structure (New)

~/charlie/
├── contacts/
│   ├── contacts.json          # Unified contact store
│   └── enrichment/            # Per-contact enrichment data
├── council/
│   ├── data/                  # Monthly data snapshots
│   └── reports/               # Council output reports
├── costs/
│   ├── daily/                 # Daily cost JSONs
│   └── dashboard/             # Static HTML dashboard
├── kb/
│   └── ingestion-log.json     # What's been ingested
├── scripts/
│   ├── cost-observatory.js    # Cost tracking
│   ├── contact-sync.js        # CRM contact sync
│   ├── meeting-prep.js        # Calendar → brief
│   ├── failover-health.sh     # Mac Mini → MBA ping
│   └── evolution-engine.js    # Drift detection
├── skills/                    # (symlink to ~/ven-skills/)
│   ├── kb-ingest/             # NEW
│   ├── crm-contacts/          # NEW
│   ├── sales-intel/           # NEW
│   ├── megan-bot/             # NEW
│   ├── ai-council/            # NEW
│   ├── cost-observatory/      # NEW
│   └── evolution-engine/      # NEW
└── .vector-store/
    └── chroma/                # ChromaDB data

End of plan. Ready for review.