Architecture blueprint · Companion to Asana synthesis

SEO Navigator on Claude Managed Agents

Adapting Asana's Managed Agents architecture to the SEO Navigator Agency OS — six department agents, seven internal workstreams, and ClickUp as the work graph.

Source talk Ara, Asana × Anthropic Adapted for SEO Navigator 2026 blueprint System of record ClickUp

Asana's pitch is that they own the work graph and Anthropic owns the outcome execution. The split scales because each side stays in its lane. The Agency OS we've been mapping is structurally the same shape — six dept agents (SEO Sentinel, Content Catalyst, Revenue Relay, Ad Arbitrage, Build Bot, PM Pulse) sitting on top of a ClickUp work graph, with Claude Skills as the curated expertise layer and Managed Agents as the execution harness. This document walks through what each Asana slide looks like once we swap their categories for ours.

01

The promise of an agentic agency

The vision Ara opened with — humans and agents collaborating across departments instead of one human prompting one agent in isolation — is the same vision that motivated the SEO Navigator skills library and the Agency OS. Six dept agents, each a curated bundle of Claude Skills, sit on top of a shared ClickUp work graph. Below them is a centaur team where every human is paired with at least one AI teammate. The arc is the agency itself.

The promise of the agentic agency An arc connecting six SEO Navigator department agents — SEO Sentinel, Content Catalyst, Revenue Relay, Ad Arbitrage, Build Bot, and PM Pulse — sitting above a row of named team members alternating with AI teammates at the base. The promise of the agentic agency Six dept agents, one shared ClickUp work graph SEO SEO SentinelRank · audits Content Content CatalystBriefs · meta QA CRM Revenue RelayGHL · follow-ups Ads Ad ArbitrageReports · copy Web/IT Build BotSchema · audits PM PM PulseReports · SOPs JN AI PA AI AT AI TR AI TG AI BN Jake Phuong Anh An Thai Trung Tung Bích Ngọc Six humans · five AI teammates · one shared graph
Fig. 01 Six dept agents on the arc above a centaur team — humans paired with AI teammates.
02

From single-player to multiplayer mode

Today, every department uses Claude in isolation. The SEO team's session doesn't know what Content just shipped. The CRM workflow has no memory of the campaign brief Ads is running. Each silo gets faster — the agency as a whole doesn't. Multiplayer mode means a shared agent layer mediated by ClickUp — task graph, RACI, and Skills memory all in one place. A correction one team makes ("Sun Stoppers' brand voice is now warmer, less technical") gets written to shared memory, and the next dept agent inherits it.

From single-player to multiplayer mode Left half shows four siloed dept teams — SEO, Content, Ads, CRM — each with one human and one isolated AI agent. Right half shows the same teams sharing a central agent hub with cross-team memory. From single-player to multiplayer mode A shared agent layer mediated by ClickUp Today — silos Tomorrow — multiplayer SEO AT AI Content PA AI Ads TG AI CRM TR AI CU AT AI PA TR AI TG Each dept's Claude is alone ClickUp graph + Skills memory shared
Fig. 02 Four isolated dept Claude sessions vs. one shared agent hub on top of ClickUp.
03

The diagnosis applies to us too

Asana's framing — that AI is failing not because the model is weak but because it's missing five pieces of context — is true at our scale, just with agency-specific surfaces. Who means which client we're working on without re-pasting context. What means whether this is a brief, a deliverable, a campaign, or a report. By when means our monthly client-reporting cycle and campaign timing. How means the agency methodology — Koray's semantic SEO, the EAV/URR triples, the CQS rubric. Why means the client's commercial OKRs and our retention pressure.

AI is context starved in our agency Five floating question phrases — who, is doing what, by when, how, and why — scattered across the canvas with thin partial curves connecting some, illustrating the fragmentation of agency context across ClickUp, Slack, GHL, Drive, and Notion. AI is context-starved in our agency Context lives in ClickUp, Slack, GHL, Drive, Notion — never assembled Who — which client doing what — brief or deliverable by when — monthly cycle how — Koray methodology and why — client OKRs
Fig. 03 Same five fragments Ara described, mapped to our agency's actual context surfaces.

The QueryMind work, the GHL CRM build, the SOP library, and the Agency OS Memory section are all closing different parts of this gap. ClickUp is the place those streams converge.

04

ClickUp as the system of action

ClickUp's Workspace → Space → Folder → List → Task → Subtask hierarchy is structurally identical to Asana's Mission → Goals → Portfolios → Projects → Tasks. The three pillars adapt directly: a superior data model lives in ClickUp custom fields (Dept, Workstream, RACI, ICE Score, Phase, Dev Flag, Automation Status, SOP Link, KPI Impacted). A multiplayer architecture lives in RACI-on-every-template plus cross-folder linking. Security and guardrails live in our human gates — CEO sign-off on hires, ad copy review before launch, OKR calibration sessions.

ClickUp as the system of action Three pillars on the left — superior data model via ClickUp custom fields, multiplayer architecture via RACI on every template, and security and guardrails via human gates — alongside ClickUp's hierarchy from workspace down through space, folder, list, task, and subtask, supported by a foundation bar of workflows, automation, integrations, resource management, KPI reporting, and RACI. ClickUp as the system of action A work graph that humans and dept agents both write into Custom fields + listsDept · RACI · ICE · Phase · Dev RACI on every templateMultiplayer — no "everyone" tasks Human gates + Dev FlagSign-off on consequential actions Workspace Spaces Folders Lists Tasks Workflows · Automation · Slack/GHL/Stripe integrations · Resource mgmt · KPI reporting · RACI
Fig. 04 Three pillars + ClickUp's hierarchy + a foundation of platform services and integrations.
05

Workstream orchestration across our six dept agents

Where Asana put cross-functional Operations/PMO workstreams on top with ICP departments below, ours has the seven internal workstreams from the 2026 blueprint as the connective layer, with the six department agents as the columns underneath. The workstreams aren't a separate system — they're the cross-functional programs that every dept agent participates in.

Workstream orchestration across the six dept agents A horizontal banner at the top labelled Internal Operations Workstreams contains seven program tiles connected by arrows: workflow, onboarding, comms, brain, templates, performance, and culture. Below sit six department agent columns — SEO Sentinel, Content Catalyst, Revenue Relay, Ad Arbitrage, Build Bot, and PM Pulse — each listing the AI-automatable tasks that agent owns. Workstream orchestration across the six dept agents Seven internal workstreams · six department agents Internal Operations · seven workstreams 01 Workflow+ SOPs 02 Onboardsimplify 03 Commsautomation 04 Brainsystem 05 Templates+ recurring 06 Perfframework 07 Culture+ mindset SEO Sentinel Rank monitoring Topical maps Tech audits AIO consensus Schema audit GSC alerts SEO reports Competitor scan 37h/wk recovered Content Catalyst Brief gen (EAV) Meta QA Citations AIO scoring Internal links Social repurpose EAV gap detect 30h/wk recovered Revenue Relay GHL follow-ups Meeting notes GHL workflows Review requests Pipeline tags SMS/email drips Lead scoring 28h/wk recovered Ad Arbitrage Perf reports Ad copy variants Competitor ads Audience maps A/B analysis Carousel feed Budget alerts 25h/wk recovered Build Bot Tech audits Schema gen Migration map Carousel feed Tech specs Code docs QA checklists 18h/wk recovered PM Pulse Monthly reports Status updates Task creation Meeting notes SOP drafts KPI dashboards Onboarding docs 27h/wk recovered
Fig. 05 Cross-functional workstream banner above six dept agent columns — 165h/wk recoverable agency-wide.

Workstream readiness for Managed Agents

Not every workstream is the same shape. Some have crisp, gradeable outputs that fit the Managed Agents harness today; others are mostly human craft that AI assists. The table below maps each workstream to its primary Claude Skill and an honest readiness call.

WS Workstream Primary Skill Phase Managed Agents fit
01 Workflow + SOP systemSOPs, workflow maps, RACI seo-navigator-agency-os W1–2 High — gradeable outputs
02 Onboarding simplificationChecklists, role packs seo-navigator-agency-os W3–6 Medium — DoD per role
03 Communication + automationSlack ↔ ClickUp routing seo-navigator-agency-os W3–7 Medium — needs Zack integration
04 Internal "brain" systemMeta Agent · digests managed-agents-doc W7–9 High — but Phase 2
05 Recurring tasks + templatesTemplate Vault seo-navigator-agency-os W1–2 High — start here
06 Performance frameworkSOWs, KPIs, OKRs seo-navigator-agency-os W3–7 Medium — depends on SOW gate
07 Culture + mindset upgradeWorkshops, coaching W4–12 Low — human-led, AI drafts only

The pattern is the same one Asana hit: the workstreams with the clearest grader rubrics (templates, SOPs, KPI scorecards) move fastest. The ones tied to human relationship and behavior (culture, sensitive client conversations) stay human-led, with AI as a drafting assistant only.

06

Why Managed Agents — for SEO Navigator

The before/after table Ara showed translates almost line-for-line into our reality. Today, every Claude Skill in the library is invoked through Claude Code or the chat interface; Jake reviews every output before it reaches a client. That pattern works at our current scale but it's the bottleneck Ara was describing — one long single-threaded conversation with the founder driving. Managed Agents replaces that with a managed harness where the rubric (our DoD) is the spec the grader iterates against.

Why Managed Agents — SEO Navigator before/after A three-row before/after comparison: implementation moves from custom Claude Code scripts per skill to pre-built Claude Skills plus the Managed Agents harness; verification moves from Jake QAing every output to agents iterating against the workstream DoD until the rubric passes; deep complex work moves from one long Claude session with Jake driving to parallel agents per workstream sharing ClickUp memory. Why Managed Agents — SEO Navigator From founder-driven workflows to outcome-driven delegation Today — Claude Code + Jake Tomorrow — Managed Agents Implementation Custom Claude Code per skill,manual orchestration Pre-built Skills + ManagedAgents harness Verification and iteration Jake QAs every outputbefore client send Agents iterate against DoDuntil the rubric passes Deep, complex work One long Claude sessionwith Jake driving Parallel dept agents,shared ClickUp memory
Fig. 06 Three layers — implementation, verification, deep work — before and after the shift.

We can focus on what's unique to us — the agency methodology, the ClickUp work graph, the client relationships — while Anthropic handles outcome quality. — Ara, Asana (paraphrased for our context)

The architecture split is the load-bearing insight. SEO Navigator owns: the ClickUp work graph (workspaces, lists, tasks, RACI, custom fields), the Skills library (Koray methodology, EAV/URR triples, CQS rubric, GHL playbooks), the GHL CRM context, the human gates, the client relationships. Managed Agents owns: multi-step task execution, verification against our DoDs, code execution for HTML deliverables, and parallel agent orchestration. Neither side does the other's job.

07

The agentic work ecosystem

The verification loop is what makes the architecture work. The grader iterates until the outcome meets the rubric — no human babysitting. For us the rubric is the DoD already written for each workstream. Faster prototyping comes from Skills as packaged expertise. Quality verification comes from running the DoD as the grader spec. The outcome evaluator closes the loop by holding the agent to the KPI scorecard.

Agentic work ecosystem for SEO Navigator A circular loop showing managed agents complete complex agency work through three connected benefits: faster prototyping powered by Claude Skills as expertise packs, quality verification where the workstream DoD becomes the rubric, and an outcome evaluator that grades against KPI scorecards and iterates until satisfied. Agentic work ecosystem Skills · DoD · KPI scorecards close the loop Complete agency deliverables Faster prototyping Skills as expertise packs Quality verification DoD becomes the rubric Outcome evaluator KPI scorecards drive iteration
Fig. 07 The closed loop, with our existing artifacts (Skills, DoDs, KPI scorecards) as the inputs.
08

Building the agency operating system

Asana's three forward bets translate to ours almost directly: multi-step workflows are the seven WS pipelines from the 2026 blueprint, team patterns are the six centaur dept agents, and proactive moves are the Meta Agent that workstream 4 (Internal Brain) is designed to deliver in Phase 2. The 12-week roadmap stays the same — stabilize W1–2, standardize W3–6, automate W7–12.

Multi-step workflows

Seven WS pipelines, end-to-end with DoD rubrics

Centaur teams

Six dept agents paired with the team — humans own WHAT and WHY

Proactive moves

Meta Agent in WS4 — wakes up when it spots blockers

SEO Navigator × Anthropic
Fig. 08 Three forward bets translated from Asana's pillars to ours.
09

Q&A — translated for our context

On the verification rubric

The good news: we already have rubrics. Every workstream in the 2026 blueprint has a written DoD ("every recurring workflow has template + SOP + video + RACI"; "deadline reminders firing automatically"; etc). Those become the grader specs without rewriting them. Where we don't have rubrics — culture, client relationships — we shouldn't try to manufacture one.

On skill maintenance

Asana keeps skills shrink-wrapped, governed centrally, with no customer authoring for now. We do too — the skills library (seo-navigator-agency-os, consensus-content-audit, google-ads-radius-optimizer, etc.) is curated and versioned. The risk to watch: as the skill count grows, we need a clear ownership map for who maintains each, who tests them, and who decides when to deprecate.

On third-party integrations

Asana said integrations happen at both the agent loop and the MCP layer. For us that means: the ClickUp MCP, the GHL MCP (the BusyBee fork or the official 36-tool endpoint), and the Slack MCP all need to be reachable from inside Managed Agents, not just from Claude Code. The integration story is symmetrical, not stacked.

On where to start

Workstream 5 (recurring tasks + templates) and workstream 1 (workflow + SOPs) are the highest-ICE places to begin — both have clear DoDs and AI-automatable outputs. Workstream 4 (Internal Brain) is the most exciting but explicitly Phase 2 (W9+). Don't skip phases.

Concrete next steps · For the next planning pass

Three decisions that pull this off the page

1. Pick the first Managed Agent prototype. The cleanest fit is workstream 5 (recurring tasks + templates) — the DoD is "Template Vault populated, 80% of execution on existing templates" — that's a measurable rubric. Build a Managed Agent whose job is "given a process description, produce a ClickUp template + SOP doc + RACI mapping that passes the DoD." Run it for 2 weeks. Compare against Jake-driven baseline.

2. Build the rubric translation layer. Every workstream's DoD currently lives in the command center HTML doc. Those need to be machine-readable rubrics that a grader can score against. This is a one-time documentation pass — turn each DoD bullet into a checklist item with an explicit pass/fail signal.

3. Figure out which dept agent gets packaged first. SEO Sentinel has the most volume (37h/wk) but Content Catalyst has the cleanest gradeable output (briefs against the EAV rubric). PM Pulse is the most universal across clients. Pick one for the v1 packaging effort and hold off on the others until we have data.

What we still don't know

Worth flagging before we commit

  • The actual Managed Agents pricing per session-hour at our throughput. Need to check current docs before sizing the budget.
  • Whether the ClickUp MCP exists today as a managed connector or whether we'd need to build/maintain a bridge. The Asana case worked because Asana built the integration; we'd need the equivalent on ClickUp's side or a self-hosted shim.
  • How much of our existing Claude Skills library can be invoked from inside a Managed Agent vs. needs to be re-authored for the harness. Some skills assume Claude Code conventions that may not transfer cleanly.
  • Whether the talk's "21+ pre-built teammates GA in March" claim holds up — taking it from Ara's word for now, would verify against Asana's actual changelog before quoting publicly.