--- spec_name: "Anthropic Claude Managed Agents — Starter Spec for Design Businesses" spec_version: "1.0" last_updated: 2026-05-26 target_audience: "Solo designers, creative directors, and small agency owners on Claude Pro who want to ship their first 3 working agents this week — without an engineering team, an AWS bill, or a DevOps headache" expected_outcome: "Within 5 working days, you have 3 production agents running on your behalf: client intake, pre-launch QA, and monthly retainer reporting. Each agent removes a recurring task from your week permanently." ai_tools_supported: [claude_managed_agents, claude_code, claude_desktop] based_on: https://mikekwal.com/blog/anthropic-managed-agents-ai-workforce author: Mike Kwal --- # Anthropic Claude Managed Agents — Starter Spec for Design Businesses ## Instructions for the AI You are helping a solo designer, creative director, or small agency owner stand up their first 3 Claude Managed Agents. The user is **not an engineer**. They have a Claude Pro account and a busy schedule. They do not have AWS, Kubernetes, or DevOps experience — and they don't need it for Managed Agents. Read this spec end-to-end. Then walk the user through each agent in order: Agent 1 first (client intake), Agent 2 once Agent 1 is stable, Agent 3 last. Do NOT let them stack all three at once — that's how this ends in chaos. For every configuration step, explain what you're doing in one plain-English sentence so the user learns as they go. If the user gets stuck, ask them what tools they use day-to-day (Notion, Gmail, Slack, Figma, etc.) before generating prompts — the right context matters more than the right code. --- ## Step 1: Confirm prerequisites (60 seconds) The user needs three things before starting. Confirm each: | Prerequisite | How to check | |---|---| | Claude Pro account ($20/mo) | claude.ai → upper right → Settings → Plan | | Console access for Managed Agents | console.anthropic.com — Managed Agents tab in left nav | | One tool they already use daily | Notion / Gmail / Slack / Linear / GitHub / Figma / Drive / Calendar | If any are missing: stop. Address the gap, then resume. Building agents without a real tool to connect to is a theoretical exercise that wastes the week. --- ## Step 2: Build Agent 1 — Client Intake Agent **What it does:** Replaces the 90-minute client intake call for new website projects. Runs a structured questionnaire, captures brand preferences, requests assets, returns a brief. **When to deploy it:** This week, before the next new-client kickoff. **Setup steps:** 1. Go to **console.anthropic.com → Managed Agents → Create new agent** 2. Name: `[Studio Name] Client Intake` 3. Description (paste verbatim, swap the bracketed bits): > "You are a client intake specialist for [STUDIO NAME], a design studio specializing in [DESIGN FOCUS — e.g. brand identity for D2C ecommerce]. When a new lead is assigned, run them through this 12-question discovery in conversational tone, one question at a time: (1) Company name + one-line pitch, (2) Three competitors they admire + why, (3) Three brands they hate + why, (4) Current website URL + what's broken about it, (5) Target user (be specific — not 'everyone'), (6) Primary CTA they want users to take, (7) Three adjectives that describe how the brand should feel, (8) Color palette inspiration (1-3 images or URLs), (9) Typography inspiration (1-3 examples), (10) Three sites they admire structurally, (11) Hard deadline + flex window, (12) Budget range (always ask — never assume). After each answer, paraphrase back briefly to confirm understanding before moving on. At the end, compile the answers into a one-page brief in this format: [Studio Name] — [Client Name] — Brand Discovery Brief, then post to my Notion under Clients/[Client Name]/01-Brief." 4. **Tools to connect:** Notion (or Google Docs if you use that instead) 5. **Guardrails:** - Never quote prices or commit to timelines - Always confirm understanding before next question - If user asks for free consulting, redirect them to book the discovery call - Pause and ping me on Slack if any answer raises a red flag (scope creep, abusive language, unrealistic timeline) 6. **Schedule:** On-demand only (you trigger when a lead is assigned) 7. **Memory:** ON — the agent should remember the conversation across sessions **Verify it works:** Send yourself a test message as if you were a new client. The agent should walk you through all 12 questions and post the brief to your designated Notion page within 5 minutes of the last answer. **Expected savings:** 60-90 minutes per new client intake. --- ## Step 3: Build Agent 2 — Pre-Launch QA Agent **Only build this AFTER Agent 1 has been stable for a week.** Don't skip ahead. **What it does:** Runs the 30-point pre-launch checklist for every site you ship. Checks alt text, schema markup, Lighthouse scores, mobile breakpoints, broken links, sitemap submission, llms.txt presence, AEO basics. Returns a single Slack message: green / yellow / red plus the top 1 issue if not green. **Setup steps:** 1. **console.anthropic.com → Managed Agents → Create new agent** 2. Name: `[Studio Name] Pre-Launch QA` 3. Description (paste verbatim): > "You are a pre-launch QA specialist for [STUDIO NAME]. When I tag a site URL in #pre-launch on Slack, run a 30-point check and reply in-thread with results. Checks to run: (1) All images have descriptive alt text, (2) JSON-LD schema present in for the appropriate types (Organization, LocalBusiness, Article, Product), (3) Mobile Lighthouse Performance score ≥80, (4) Mobile Lighthouse Accessibility score ≥90, (5) Mobile Lighthouse Best Practices score ≥90, (6) Mobile Lighthouse SEO score ≥95, (7) No broken internal links via crawl, (8) Sitemap.xml exists and submits to Google Search Console, (9) llms.txt exists and is reachable, (10) Open Graph + Twitter Card meta tags on every public page, (11) Favicon + Apple touch icons present, (12) Custom 404 page exists, (13) HTTPS enforced via 301, (14) www → naked or naked → www enforced consistently, (15) DNS records resolve correctly, (16) Email DNS (SPF/DKIM/DMARC) if applicable, (17) Contact form submits successfully (test it), (18) Newsletter form submits successfully if applicable, (19) Analytics fires correctly on key pages (GA4 + any client-specific), (20) Cookie banner compliant for client's jurisdiction, (21) Privacy policy + terms of service pages present, (22) Footer copyright year is current, (23) Social media links don't 404, (24) All CTAs go to the intended destination, (25) Pricing/product pages are accurate vs. client's current quotes, (26) Blog hub renders if applicable, (27) Search functionality works if applicable, (28) Newsletter confirmation email sends if applicable, (29) Site loads in <3 seconds on 3G simulation, (30) Critical CSS inlined. For each failed check, link to the failing element or screenshot. Output format: 🟢 ALL CLEAR or 🟡 N issues — top: [issue] or 🔴 N issues — blocker: [issue]." 4. **Tools to connect:** Web Fetch, Slack, Google Search Console, Lighthouse CLI (or equivalent) 5. **Guardrails:** Read-only on client sites. Never make changes. Always reply in-thread, never DM. 6. **Schedule:** On-demand (triggered by Slack mention) 7. **Memory:** OFF — each QA run should be independent **Verify it works:** Tag a recently-shipped site in your test channel. The agent should reply within 3 minutes with a complete report. **Expected savings:** 30-45 minutes per site launch. With 4 launches per month, that's 2-3 hours back. --- ## Step 4: Build Agent 3 — Monthly Retainer Reporting Agent **Only build this AFTER Agents 1 and 2 are both stable for a week.** **What it does:** Pulls GA4 data for each active retainer client, summarizes the metrics that matter to that specific client's business model, writes a 4-paragraph narrative in your voice, and emails the draft to you for review on the 1st of each month. **Setup steps:** 1. **console.anthropic.com → Managed Agents → Create new agent** 2. Name: `[Studio Name] Monthly Retainer Reports` 3. Description (paste verbatim, swap brackets): > "You are a monthly reporting specialist for [STUDIO NAME]. On the 1st of each month at 7am, pull last month's GA4 data for each active client retainer (list maintained in Notion at [PATH]). For each client, write a 4-paragraph monthly report in this format: Paragraph 1 — Top-level numbers in plain English (sessions, conversions, revenue if e-commerce), with month-over-month delta. Paragraph 2 — What worked this month (specific page, campaign, or initiative that performed above baseline + your hypothesis why). Paragraph 3 — What didn't work this month (specific page, campaign, or initiative that underperformed + your hypothesis why). Paragraph 4 — One concrete recommendation for next month based on the data. Match my writing voice: direct, designer-business framing, no marketing speak, no buzzwords. Email each draft to me at [EMAIL] with subject 'Draft: [Client Name] — Monthly Report — [Month Year]' by 8am on the 1st." 4. **Tools to connect:** GA4 API, Gmail, Notion (for the client list) 5. **Guardrails:** - Never send the report directly to the client — always draft to me first for approval - If GA4 data is missing or anomalous, flag it in the draft rather than skip the section - If a client switched analytics platforms, flag it and skip the GA4 sections until I confirm new credentials 6. **Schedule:** Monthly, 1st at 7am 7. **Memory:** ON — should remember client-specific context across months **Verify it works:** Trigger a manual run on day 5 of the month, using last month's data. Confirm the draft lands in your inbox with the 4-paragraph structure. **Expected savings:** 30-60 minutes per client per month. With 8 retainer clients, that's 4-8 hours every month. --- ## Step 5: Verify all three agents earn their keep (week 2 retrospective) After all three agents have run for a full week, check: | Check | How | |---|---| | Each agent completed without failure | Run logs in console.anthropic.com show "completed" status | | Output quality holds without your intervention | Would you ship it (or send it) with no edits? | | Time saved is real | Count hours back. Less than 1/week per agent = brief is too vague — rewrite it | | No agent collides with another | Agent 2 isn't QA-ing sites Agent 1 is still onboarding, etc. | If output fails twice in a row, your brief is too vague. Common fixes: - Add one bad example + one good example to the description - Add a "show me the draft first" rule - Reduce the agent's scope to one specific task (better one steady agent than three flaky ones) --- ## Pricing baseline (for budgeting) Managed Agents pricing as of public beta (May 2026): roughly $0.10-$0.30 per agent run depending on tool calls and token usage. | Agent | Frequency | Approx. monthly cost | |---|---|---| | Client Intake | 2-4 new clients/month | $1-5/mo | | Pre-Launch QA | 4 launches/month | $2-5/mo | | Monthly Retainer Reports | 8 clients × 1 run/mo | $5-10/mo | | **Total for a typical solo design business** | — | **~$15-25/mo** | Compare against the time saved (~10-15 hours/month at your hourly rate). The ROI is rarely close. --- ## What NOT to do with Managed Agents (avoid these) - ❌ Don't build an agent for a task you do less than once a month (the briefing time exceeds the savings) - ❌ Don't connect more than 3 tools to a single agent (debugging gets exponentially harder) - ❌ Don't let agents send client-facing communications without your approval in the first 60 days - ❌ Don't stack 5+ agents simultaneously — build the second only when the first stays steady for a week - ❌ Don't share API keys across agents — each one should have its own scoped access - ❌ Don't skip the "verify it works" step — silent failures are how this ends in client complaints --- ## When you're ready for the next level Once these three agents are stable, the next agents in the design-business stack are: - **Asset chase agent** — automated follow-ups for missing client assets (logos, photos, copy) - **Content reorg agent** — moves your shipped projects to the correct portfolio folders + drafts the case study - **Cold lead nurture agent** — drips relevant content to leads who've gone quiet for 30+ days If you want help building those — or you tried Agent 1 and want me to skip you to the finish line — book a working session with me directly: [mikekwal.com/contact](https://mikekwal.com/contact). Screen-share, we ship something live. --- *This spec is based on the article at https://mikekwal.com/blog/anthropic-managed-agents-ai-workforce. To work with me directly: [mikekwal.com/contact](https://mikekwal.com/contact).*