A 30-day deployment of Layer 2 — the Technical Architecture. Phase 1 proved the manual model; Phase 2 makes it autonomous. Stand up the reusable orchestration harness, ship the 6-agent Managed-Agent swarm, wire the Unified Command Center, then productize it all as the Hybrid OS retainer.
Four things must be true before a single agent runs. Confirm these in the Day-0 kickoff so Week 1 isn't spent chasing access.
Anthropic Managed Agents beta access confirmed, beta header set, and the $0.08/session-hour pricing model understood. Sprint 1 test budget ~$50–100 approved.
SEO Utils MCP reachable; external API keys ready (Apify, Firecrawl, Google Maps, OpenAI, Gemini, Perplexity); ClickUp status trigger + Slack slash-command path scoped.
Hung owns AI automation (harness, agents, n8n). Trung (IT) owns infrastructure + website/Astro. Phanh owns Sentinel's system prompt + rubric. Jake approves spend + the architecture map. PM owns the ClickUp Agent Registry.
Phase 1 must be green: CRM live, repeatable onboarding, monthly reporting working. Layer 2 automates these proven manual processes — don't automate what isn't yet proven.
Building the agent fleet before the manual model is proven is the classic premature-optimization trap. Phase 1 earned the right to automate; Phase 2 cashes it in. Readiness checks keep the order honest.
Build the reusable orchestration harness once, then deploy SEO Sentinel v1 as the proof agent. This is the hard week — the abstraction built here is what makes agents 2–6 trivial.
| Task | Owner | Hrs | Source | |
|---|---|---|---|---|
| Build the dispatcher + config-file pattern (./agents/<name>.json: agent_id, kickoff prompt, ClickUp fields, delivery channels) | Hung | 6h | Reuse PRD — Orchestration | |
| Implement session lifecycle state machine + SSE event handling with reconnect logic | Hung | 5h | Reuse PRD → lifecycle | |
| Add idempotency, logging, and the HITL gate handler (stubbed for v1 — goes live with PM Pulse in Week 3) | Hung | 4h | Reuse PRD → HITL | |
| Deploy to VPS and pass the 7-test integration suite | Hung | 3h | Reuse PRD → tests | |
| Approve Anthropic API spend budget (~$50–100 Sprint 1) + sign off on the harness architecture | Jake | 1h | PRD Local Automation |
The harness is the leverage. Built properly once for Sentinel, it carries Catalyst, Revenue Relay, Ad Arbitrage, Build Bot, and PM Pulse for free. The abstraction pays for itself on the second agent — classic “better beats more.”
| Task | Owner | Hrs | Source | |
|---|---|---|---|---|
| Create the SEO Sentinel agent + environment config; wire 5 modules (GBP, On-Page, Geographic grid, Citations, AI Visibility) via bash + SEO Utils MCP | Hung | 5h | PRD Local Automation → §7 | |
| Phanh writes the v1 system prompt + scoring rubric (coverage, accuracy, structure) | Phanh | 3h | SEO Lead Handoff | |
| Wire triggers: ClickUp status → “Ready” and Slack slash-command; output posts to Slack + updates the ClickUp task | Hung | 3h | PRD → triggers | |
| Run T1 + T2 dry runs on a synthetic client; iterate the system prompt with the Phanh | Phanh | 2h | PRD → testing |
Sentinel is read-only on purpose: the riskiest thing an agent can do is publish. Prove accuracy on audits first, earn trust, then graduate to client-facing actions behind HITL gates. ~$1.75/run vs. ~3 analyst hours is the unit win that funds the whole fleet.
Cash in the harness abstraction: plug in Content Catalyst and Revenue Relay via config files, no new orchestration code. In parallel, stand up the n8n playbook as the no-Managed-Agent fallback path.
| Task | Owner | Hrs | Source | |
|---|---|---|---|---|
| Drop in catalyst.json — wire Content Catalyst to QueryMind MCP (brief gen, meta optimization, AEO scoring) | Hung | 3h | Reuse PRD → config pattern | |
| Content Lead writes Catalyst's system prompt + rubric; validate brief output vs. a reference brief | Nhu | 3h | QueryMind pipelines | |
| Drop in revenue-relay.json — wire Revenue Relay to GoHighLevel MCP (follow-up sequences, lead scoring) | Hung | 3h | SNMS CRM + Reuse PRD | |
| Confirm the zero-code claim: log the diff — only config files added, dispatcher untouched | Jake | 1h | Reuse PRD → acceptance |
Agent #1 took a week. Agents #2 and #3 take a config file each. That curve is the entire thesis of Layer 2 — build the system, then the marginal cost of capacity collapses toward zero.
| Task | Owner | Hrs | Source | |
|---|---|---|---|---|
| Build the 4 n8n modules: GBP Intelligence (44-point), On-Page Engine, Geographic grid (9/25-pt via Apify), Citation/NAP | Hung | 6h | Local SEO Automation Playbook | |
| Wire the pipeline: web form → n8n parallel APIs → Code scoring → OpenAI HTML report → SendGrid → Airtable | Hung | 4h | n8n Playbook → flow | |
| Phanh validates n8n report output parity against a Sentinel run on the same client | Phanh | 2h | n8n Playbook |
Two paths to the same outcome (Managed Agent + n8n) is risk reversal at the infrastructure level. If Anthropic pricing or features shift, delivery doesn't stop — the n8n path keeps the SEO audits flowing.
Complete the fleet (Ad Arbitrage, Build Bot), then deploy PM Pulse — the coordinator that delegates to specialists and activates the HITL gates for client-facing output. Wire it all into the Unified Command Center.
| Task | Owner | Hrs | Source | |
|---|---|---|---|---|
| Deploy Ad Arbitrage (ad-arbitrage.json) — Google + Meta reporting, copy variants, audit reports; Sang/Tung write its rubric | Sang + Tung | 3h | Manage Agent SEO swarm | |
| Deploy Build Bot (build-bot.json) — migrations, schema audits, Lighthouse + Playwright UAT | Hung | 3h | Manage Agent SEO | |
| Deploy PM Pulse coordinator — multi-agent delegation, output synthesis, Slack posting | Hung | 4h | AI Orchestration diagram | |
| Activate + test the HITL Slack gate on a client-facing deliverable (email draft / GBP post) — nothing ships without approval | Jake | 2h | Reuse PRD → HITL |
PM Pulse is where agents start producing client-facing work, so the HITL gate goes live here. The centaur model holds: AI throughput, human approval before anything reaches a client. Speed without a quality gate is just faster mistakes.
| Task | Owner | Hrs | Source | |
|---|---|---|---|---|
| Stand up the Unified Command Center (ClickUp + Slack) across the 7 workstreams | Hung | 4h | Automating Agency Workflows | |
| Build the ClickUp Agent Registry v1 — agent_id, owner, scope, cost/run, last run, status per agent | PM | 3h | SN Workstream PM | |
| Define governance: spend caps per agent, escalation paths, and the weekly agent-fleet review cadence | PM | 2h | SN Workstream PM → governance |
The Command Center turns six invisible cloud agents into a dashboard a human can run. Cost-per-run, accuracy, and hours saved are visible per agent — the basis for the Phase 2 graduation metrics and every optimization that follows.
Turn the working swarm into a sellable product, enroll the first 2 beta clients, and prove the graduation metrics: SOP accuracy ≥90%, fleet accuracy ≥85%, and ≥30% hour reduction.
| Task | Owner | Hrs | Source | |
|---|---|---|---|---|
| Package the Hybrid OS offer — scope, deliverables, price anchor vs. Detailers Roadmap, guarantee | Jake | 3h | AI driven agency → Hybrid OS | |
| Build the Hybrid OS GHL snapshot (the productized delivery container) | Trung Le | 5h | 90-D Transfo → G4 Hybrid OS | |
| Enroll 2 beta clients — one call each, grandfather pricing in exchange for feedback | Jake | 2h | 90-D Transfo → beta |
A second premium product to the same client base is the highest-leverage LTV:CAC move there is. The swarm already runs delivery, so gross margin on the Hybrid OS is structurally higher — that's the “triple the margin” claim made real.
| Task | Owner | Hrs | Source | |
|---|---|---|---|---|
| Run the 12-agent parallel UAT (239 of 334 checks automated, ~40 min, ~$9.65/run) on both beta clients | Trung Le | 3h | AI driven agency → UAT | |
| Measure the bar: SOP accuracy ≥90%, fleet accuracy ≥85%, ≥30% hour reduction vs. manual | Jake | 2h | 90-D Transfo → criteria | |
| Compile the Phase 2 retrospective — what to optimize (Stage 6) and Phase 3 (Layer 1) input | Jake | 2h | Master Index → Layer 2 review |
Six Managed Agents on Anthropic infrastructure — one coordinator and five specialists — deployed in 8 days versus four weeks of DIY. Hours saved are weekly capacity recovered across the client book. Pricing is standard tokens + $0.08/session-hour, no idle charges.
Delegates to specialists, synthesizes their outputs, posts to Slack. Owns the HITL gate for client-facing deliverables.
Reports, audits, rank monitoring, AEO. Runs the 5 Local SEO modules at ~$1.75/run, replacing ~3h of analyst work.
Brief generation, meta optimization, AEO scoring — wired to QueryMind MCP.
GHL automation, follow-up sequences, lead scoring — wired to the GoHighLevel MCP.
Google + Meta reporting, ad-copy variants, audit reports across both ad platforms.
Site migrations, schema audits, Lighthouse + Playwright UAT runs.
The end-to-end chain for Agent #1, straight from the Master Index's Layer 2 pipeline. Each link is a real document — WHAT the agent does, HOW it's triggered, WHO owns the deliverables, and the n8n alternative path.
PRD Local Automation — 19-section spec, 5 modules
Reuse PRD — config pattern, SSE, HITL
SEO Lead Handoff — prompt + rubric
Local SEO Automation Playbook — 4 modules
Live payback and return for the fleet. Defaults are grounded in the source docs — SEO Sentinel replaces ~3 analyst hours at ~$1.75/run. Toggle to the fleet view to see total weekly capacity recovered in dollars.
The dollar return is large, but the point isn't token savings — it's that ~165 hours/week of human capacity move from execution to judgment, relationships, and selling the Hybrid OS. That reallocation is what triples margin, not the $1.75 line item.
Phase 2 is the AI-automation buildout — Hung is Responsible for the technical spine (agents, harness, n8n); Trung (IT) owns infrastructure, deployment, and website/Astro builds. Jake stays Accountable and owns the two judgment calls that can't be delegated: spend approval and the HITL gate on client-facing output.
| Workstream | Jake | Hung (AI) | Phanh | Content (Nhu) | Ads (Sang/Tung) | PM |
|---|---|---|---|---|---|---|
| Orchestration harness | A | R | I | I | I | C |
| SEO Sentinel v1 | A | R | R | I | I | I |
| Fleet graduation (agents 2–5) | A | R | C | C | R | I |
| PM Pulse + HITL gates | R | R | C | C | C | A |
| Command Center + Registry | I | R | I | I | I | A |
| Hybrid OS productization | R | C | I | I | I | C |
Phase 2's risk profile is higher than Phase 1's — you're now dependent on a beta platform and shipping AI-produced work toward clients. Each risk has a mitigation already wired into the plan.
The fleet runs on Managed Agents (public beta). A pricing change, feature deprecation, or outage would hit delivery directly — this is the top risk named in the strategic vision.
Mitigation: The n8n playbook (Week 2) delivers the same SEO audits without Managed Agents. Skill definitions are portable to other frontier models; all client data lives in GHL / WordPress / ClickUp, never on Anthropic infrastructure.
Once PM Pulse produces email drafts or GBP posts, an unreviewed error reaches a client and damages trust.
Mitigation: HITL Slack gate is mandatory for any client-facing action (Week 3). Sentinel and the early agents are read-only by design — trust is earned on audits before agents touch published output.
Every agent plugs into one orchestration layer. A harness bug or VPS outage takes the whole fleet down at once.
Mitigation: 7-test integration suite gates the harness before any agent rides on it (Gate A). Idempotency + reconnect logic built in; the n8n path is an independent fallback for the highest-volume workflow.
Agents run, but outputs aren't reliable enough to graduate — the Phase 2 metric fails and beta clients get sub-par work.
Mitigation: Per-agent rubrics + side-by-side validation against reference outputs at each gate. Gate D blocks scaling until ≥85% fleet accuracy and ≥30% hour reduction are measured, not assumed.
Runaway runs or verbose prompts inflate spend beyond the ~$1.75/run economics.
Mitigation: $0.08/session-hour with no idle charges keeps the floor low. The Agent Registry sets per-agent spend caps and the Command Center surfaces cost-per-run weekly. Sprint 1 budget is a contained ~$50–100.
Every sprint traces to specific Layer 2 documents in the 1st Batch. Links resolve to the real files in this folder.
| Sprint | Primary source documents | Theme |
|---|---|---|
| S-1A | Reuse PRD — Agent Orchestration, PRD Local Automation | Managed Agent Arch |
| S-1B | PRD Local Automation, SEO Lead Handoff | Sentinel Pipeline |
| S-2A | Reuse PRD (config pattern), QueryMind, SNMS CRM | Managed Agent Arch |
| S-2B | Local SEO Automation Playbook (n8n) | Workflow Automation |
| S-3A | Manage Agent SEO (swarm), AI Orchestration, Reuse PRD (HITL) | Managed Agent Arch |
| S-3B | Automating Agency Workflows, SN Workstream PM | Workflow Automation |
| S-4A | AI driven agency (Hybrid OS), 90-D Transfo (G4) | Productization |
| S-4B | AI driven agency (UAT), Managed Agent — Monthly Report | Productization |
With delivery automated and the Hybrid OS proven on betas, Phase 3 deploys Layer 1 (Strategic North Star): positioning, the outbound → content flywheel, and the hiring plan — then scales beta clients to paying.