WF4 MVP · v1.0
Monthly Report Automation
The first Claude Managed Agent deployment. Ship one agent to production on 3 real clients before opening a second front. Every subsequent workflow re-uses the infrastructure built here — environment, orchestration, secrets, HITL gates, and MCP integrations. Getting these right once is 80% of the work for every future agent.
📍 Start Here
Dependency Map
Task graph + external requirements · 18 starter tasks · 28 cross-owner handoffs · 37 external items to confirm
Open map →
▸ Team — Click any card to open that owner's dashboard
Each dashboard has Input → Output flow, task cards with Prerequisites + Steps + Done-when, and handoff map.
▸ The 4-Workflow Sequence
Each workflow gates on the prior one. Do not start the next until all promotion criteria are green.
Workflow 4 — Monthly Report Automation · YOU ARE HERE
Agent: SEO Sentinel v1 (WF4 scope only) · Gate: 3 clients × 1 cycle autonomous, <5% grader fail, <1hr review/run. Phuong Anh signs off.
Workflow 1 — Local SEO Automation (5-module audit)
Agent: SEO Sentinel v2 (adds WF1 modules) · Gate: 3 clients onboarded end-to-end. No WF4 regressions.
Workflow 2 — QueryMind Pipelines
Agent: Content Catalyst (new) · Gate: QueryMind MCP or CLI exists (Jake's J9 decision). Catalyst validates topical map on 1 client.
Workflow 3 — SEO NEO Campaign (5-agent pipeline)
Agent: SEO NEO pipeline (separate sub-agent system) · Gate: Runs after WF1/WF2 stable. Requires Playwright wrapper or VA workflow.
🔑 Critical dependency — QueryMind MCP: Workflow 2 needs a QueryMind MCP server or CLI. QueryMind today has no MCP endpoint. Options: (a) Trung builds a thin MCP wrapper (~3–5 days), (b) wrap the QueryMind CLI if one exists, (c) defer WF2. Jake decides via task J9 before WF4 ships so Trung can scope the work into the WF1 → WF2 gap. Do NOT start WF2 work until WF4 is in production.
▸ Critical Path
The tasks that gate everything else. Slip any of these and the ship date moves.
- Week 1: Jake J2 (budget) + J6 (PRD sign-off) · Trung T1 (API pre-flight) + T2 (tunnel stabilization) · Phuong Anh P1 (template)
- Week 2: An Thai A1 (KPI glossary, 5h) + A3 (query map) · Trung T3 (environment) + T5 (upload mounts) + T6 (create agent)
- Week 3: Trung T11 (T1 smoke test) → Phuong Anh P5 validates → Trung T12 (T2 real-client) → Phuong Anh P6 → Jake J8 go/no-go
- Week 4+: Production rollout · Linh L3 per-cycle reviews · Phuong Anh P7 month-1 QA · Jake J10 cost review
▸ Open Questions
What we don't know yet. Flag these before T2.
7 Questions for Anthropic (Trung asks support)
#
Question
Why it matters
Fallback
Q1
Secrets mechanism — native env_vars in Environment schema, or pass via handoff?
Plan A (clean) vs Plan B (keys in logs)
Plan B + restrict Console access; Plan A migration in v1.1
Q2
Billing during MCP waits — does session-hour accrue while waiting on MCP response?
Cost model accuracy. SEO Utils tunnel: 2-5s per call
Measure on T1 + update projections
Q3
Max session duration — hard kill at 2h / 6h / 24h?
WF4 <30min is well under any cap, but confirm for multi-client future
Assume 6h cap
Q4
Container CPU/RAM/disk — per-session limits?
Pandas on long-tenure client data could be memory-heavy
Stress test on T1
Q5
Data residency — inference_geo applicability to Managed Agents?
Future non-US client compliance
No current clients need this
Q6
Skill file mount flow — confirm end-to-end upload → agent-config → container-path
Entire build depends on this
Test with one file first (T5 early)
Q7
Research Preview pricing at GA — callable_agents, memory, outcomes
Client pricing decisions downstream
20% margin cushion + quarterly review
6 Internal Questions (answer before T2)
#
Question
Owner
Why
Q8
Actual token count after mounts + system prompt. If >80K, what do we trim?
Trung measures; Phuong Anh + An Thai decide trim
Cost + performance
Q9
SEO Utils tunnel uptime over last 30 days. If <99%, HA plan needed.
Trung
Reliability
Q10
Is local-seo-city-page-auditor redundant with koray-city-page-auditor?
An Thai
Skill inventory hygiene
Q11
GHL contact custom field for embed — exists on all client records or needs creation?
Trung + Jake
Blocks T9
Q12
DOCX output required, or is Markdown + PDF-on-demand sufficient?
Phuong Anh (client-facing preference)
Environment package decision
Q13
Price same and deliver faster, or use speed-up for capacity expansion?
Jake
Business model — not technical
▸ MVP Ship Criteria
v1 ships when ALL of the following are true. Jake signs off before build; amendments require re-sign.
- Sentinel v1 agent exists —
agent_id in version control
- env_seonav_prod environment created, packages installed, mount files uploaded
- Secrets plan decided (Plan A or Plan B) and documented in ClickUp
- Orchestration script runs on VPS — cron + ClickUp webhook triggers. Posts to Drive + Slack + ClickUp.
- T1 passes 3 consecutive clean runs on synthetic client
- T2 passes on real client. Phuong Anh validates; Jake signs off
- Cost measured — average <$3/run (2x buffer over projection)
- Runbook in ClickUp — retrigger, log inspection, secret rotation, escalation, rollback
- Known issues + workarounds logged for v1.1 planning
- Monitoring active — daily cost alert, per-session failure alert
▸ Promotion Gate → WF1
Do not start WF1 work (Sentinel v2 / Local SEO Automation) until ALL green.
- WF4 in production for 30+ days
- 3 concurrent clients running WF4 cleanly
- Grader failure rate <5% (Phuong Anh month-end call)
- Team reports LESS time on monthly reports than before, not more
- Cost model validated — actual within 50% of projection
- At least one systematic improvement deployed (rubric v1.1 from month 1 learnings)