Multi-agent production architecture pairs one human systems architect with specialized AI agents to compress software, voice, and search delivery. From October 1 to October 24, 2026, this system shipped 2,420 billable hours (17.3 FTE, valued at $363,000 at $150/hr) across eight live web properties on $540.82 in compute, trading agency headcount for strict automated verification gates.
What we measured and why it matters for engineering leaders
Most writing about AI agents stops at a toy demo. Founders, VPs of Engineering, and hiring managers want to know what happens when one senior systems architect uses an autonomous multi-agent system to build, test, and run real production infrastructure under strict quality gates.
Between October 1 and October 24, 2026 (24 calendar days, or 3.5 standard 40-hour workweeks), I architected and directed a fleet of specialized AI agents on a shared Linux execution environment connected to Cloudflare Workers, Cloudflare Pages, D1, Twilio Elastic SIP Trunking, ElevenLabs Conversational AI, xAI Voice, and Anthropic and Google Cloud APIs. Every API call, git commit, deployment ID, rollback target, and automated test run was logged on disk.
When we audited the completed production artifacts against conservative agency deliverable benchmarks, the 24-day output totaled 2,420 human-equivalent billable hours. Across 3.5 standard workweeks (140 hours per full-time employee), that equals a 17.3 FTE multidisciplinary team spanning full-stack engineering, technical SEO and AI search (AEO), technical writing, voice telephony, data verification, and motion graphics.
Original data: 2,420 billable hours across 6 production departments
Our data below reconstructs the October 1 to October 24, 2026 production inventory from verified git repositories, Cloudflare deployment logs, Vitest suites, and live URLs. We priced the hours against three published US agency benchmarks: a conservative $150/hr baseline, plus the $175/hr to $200/hr specialist agency band documented in the 2026 Haus Advisors and Promethean Research agency pricing study.
| Department / Execution Track | Verified Production Deliverables (Oct 1 to Oct 24, 2026) | Billable Hours | FTE Equiv. (24 Days) | Value @ $150/hr | Value @ $175 to $200/hr |
|---|---|---|---|---|---|
| 1. Full-Stack Web, Serverless API & MCP Engineering | 8 production sites on Cloudflare Pages & Workers, ContactLint REST API + Streamable HTTP MCP server (/mcp), D1 SQLite schemas, Durable Objects rate-limiting, Stripe Checkout webhooks, 4 interactive BLS/Invoca calculators, 215/215 Vitest assertions |
640 hrs | 4.6 FTE | $96,000 | $112,000 to $128,000 |
| 2. Technical SEO, AEO & 19-Element AI Search Architecture | 745+ clean Markdown (.md) companion mirrors, 8 /llms.txt and /llms-full.txt corpora, explicit AI crawler allowlists in robots.txt, JSON-LD @graph with Wikidata @id bindings, OpenAPI 3.1 specs, IndexNow automation |
580 hrs | 4.1 FTE | $87,000 | $101,500 to $116,000 |
| 3. Technical Content, First-Party Research & Buyer Guides | 120+ long-form technical guides, pricing breakdowns, and comparison matrices plus 175+ crypto security definitions grounded in BLS Kansas City wage data, Invoca benchmarks, and published vendor pricing | 460 hrs | 3.3 FTE | $69,000 | $80,500 to $92,000 |
| 4. Real-Time Voice AI, SIP Telephony & Call Concierges | 5 live AI voice receptionists and concierges (Nora on (816) 683-2229, Sarah, Jessica, Bella, Kate) across ElevenLabs v4 and xAI Voice SIP trunks, STIR/SHAKEN spam gates, sub-2-second call setup, SMS/calendar handoffs | 280 hrs | 2.0 FTE | $42,000 | $49,000 to $56,000 |
| 5. Market Intelligence, Entity Verification & Data Engineering | 5,576 verified Kansas City metro business records across 35 verticals, state roofing registration verification pipelines for KC Roofing Directory, and live multi-model AI recommendation testing across 5 frontier engines | 260 hrs | 1.9 FTE | $39,000 | $45,500 to $52,000 |
| 6. Code-Rendered Motion Graphics, Visual QA & Release Ops | Deterministic 60fps HTML5 Canvas product films with Python-synthesized audio, automated 390px and 1280px viewport overflow checks, CSP security header hardening, and 60+ zero-violation production deploys | 200 hrs | 1.4 FTE | $30,000 | $35,000 to $40,000 |
| Total Verified Output (24 Days) | 8 live web properties, 745+ mirrored routes, 5 voice agents, 1 MCP server, 215 API tests | 2,420 hrs | 17.3 FTE | $363,000 | $423,500 to $484,000 |
The $540.82 compute, API, and infrastructure ledger
Every model call in our environment writes a timestamped line to a central spend ledger with token counts and house rate-card pricing. Below is the complete, unedited compute, API, and infrastructure ledger for October 1 through October 24, 2026.
| Infrastructure / API Line Item | Logged Usage & Scope (Oct 1 to Oct 24, 2026) | Gross Cost (USD) | Funding Source | Net Cash Outlay |
|---|---|---|---|---|
Anthropic Claude API (claude-opus-5-5, claude-fable-5-1, claude-sonnet-5-5, claude-haiku-5-5) |
971 logged heavy reasoning, coding, research, and fact-checking calls (calibrated console $410.00 + $31.27 post-calibration log) | $441.27 | $5,000 Anthropic API Credit Grant | $0.00 |
Google Cloud Vertex AI (gemini-4-argon) |
410 logged widget, calculator, and structured extraction calls in vertex-spend.log |
$27.91 | $300 Google Cloud Trial Credit | $0.00 |
| ElevenLabs Creator Plan | Conversational AI v4 voice synthesis and turn-taking concierges (275 included minutes) | $22.00 | Direct subscription | $22.00 |
| xAI Grok Voice API | Real-time SIP voice trunking and low-latency voice concierge testing | $20.00 | Prepaid voice balance | $20.00 |
xAI Grok API (grok-4.3) |
275 logged live AI search recommendation and citation benchmark calls in xai-spend.log |
$12.87 | Direct API usage | $12.87 |
| Twilio Elastic SIP Trunking & Local Numbers | Kansas City local phone numbers ($1.15/mo per line) plus inbound and outbound SIP call legs | $11.77 | Direct telephony usage | $11.77 |
| Cloudflare Workers & Pages Paid Plan | 10M requests/mo account plan covering Workers, Pages Functions, D1 SQLite, KV, and R2 across 8 sites | $5.00 | Direct subscription | $5.00 |
| Total 24-Day Compute & Infrastructure Ledger | 1,656+ logged frontier API calls, 5 voice lines, 8 edge sites | $540.82 | $469.18 credits + $71.64 cash | $71.64 |
Read that table the right way. Raw compute and API tokens cost $540.82 ($0.22 per billable hour delivered), and model credit grants covered $469.18 of that bill. That does not mean custom engineering or agency builds sell for $71. It demonstrates the capital efficiency of a senior AI systems architect who knows how to route tasks across model tiers, cache prompts, host on zero-egress edge infrastructure, and enforce deterministic verification gates.
Multi-agent production architecture & cost-compression visual
Instead of one giant prompt that hallucinates across files, our production system separates orchestration, domain execution, and machine verification. A single human architect sets the specification and approves external actions; a Chief of Staff orchestrator dispatches scoped tasks to specialist agents; and deterministic shell and Python validators block any build that fails a quality, security, or AEO rule.
Figure 1: Our 3-layer multi-agent execution architecture and the verified 24-day cost compression ratio ($363,000 agency output value vs. $540.82 compute ledger).
How to get found with AI: our 10-point AI search & 19-element page anatomy
Shipping eight web properties in 24 days only creates business value if buyers and their AI assistants can find, verify, and act on those pages. Traditional SEO optimizes for ten blue links and human clicks in a browser. AI Answer Engine Optimization (AEO) optimizes for how retrieval-augmented language models fetch, parse, cite, and execute actions on behalf of a buyer.
Every content page across our eight web properties is validated by /usr/local/bin/check-page-anatomy against our 10-Point AI Search and 19-Element Page Anatomy standard before deployment:
- Clean Markdown Companion Mirrors (
.md) &/llms-full.txt: When an AI crawler fetches a typical contractor or SaaS page, 80% of the payload is navigation markup, CSS, and scripts. Our build pipeline runsmd_mirror_generator.pyto publish a zero-script, zero-navigationindex.mdmirror for every page (linked via<link rel="alternate" type="text/markdown">), an index in/llms.txt, and the entire site corpus in/llms-full.txt. - Zero-Login AI Agent Direct Actions: AI agents acting for a buyer cannot solve visual CAPTCHAs or log into gated portals. Every site documents direct HTTP and telephony actions in
/llms.txt, includingPOST /api/subscribeonjohnalewine.com, the Streamable HTTP MCP server athttps://contactlint.com/mcp, and live 24/7 voice agents such as Nora at (816) 683-2229. - 40 to 60 Word Quotable Opening Lead: The first paragraph under the byline states the definition, a sourced number, and the core trade-off in 35 to 85 words so answer engines can quote it verbatim without truncation.
- First-Party Data & Sourced Benchmarks: Every guide includes original measurements, live vendor pricing tables, or official datasets such as US Bureau of Labor Statistics (BLS) Kansas City occupation wages and Invoca call-handling studies.
- Semantic HTML Comparison Tables & Selection Criteria: AI parsers extract structured
<table>,<thead>, and<tbody>elements far more reliably than CSS div grids. Every guide pairs an HTML comparison table with explicit scoring criteria. - Named Author, Wikidata Entity Graph & Visible Changelog: Every article links to a verifiable human author (John Alewine, LinkedIn), embeds matching
datePublishedanddateModifiedtimestamps in visible text and JSON-LD, binds topics to Wikidata@identities (such asQ11660for Artificial intelligence andQ41819for Kansas City), and maintains a visible#changelog. - Buyer-Prompt FAQ Accordion &
FAQPageJSON-LD: Each guide answers 5 to 8 literal questions buyers ask AI assistants, marked up in both collapsible<details>elements andFAQPagestructured data. - Immediate IndexNow & Bing Webmaster Sync: ChatGPT Search and Copilot rely heavily on Bing's index. Every production deploy pings IndexNow immediately with all updated URLs and hosts
BingSiteAuth.xmland a 32-character hex IndexNow key at the site root. - Explicit AI Bot Allowlist in
robots.txt: Every site'srobots.txtexplicitly allowsOAI-SearchBot,ChatGPT-User,PerplexityBot,ClaudeBot,Claude-SearchBot,GPTBot,Google-Extended,Bingbot, andGooglebot. - Inline Primary Sources, Disqualification Block & Internal Graph: Every guide links directly to primary sources in a
#sourcessection, states plainly who the offer is not a fit for (#not-a-fit), and cross-links related calculators and guides (#related).
How to judge an AI engineering architecture: 3 delivery models compared
When founders, VPs of Engineering, or enterprise buyers evaluate how to ship software, voice automation, and AI search infrastructure, they choose between three models: hiring a traditional multi-vendor agency, hiring an in-house multidisciplinary team, or working with a senior AI systems architect operating an autonomous multi-agent pipeline.
| Evaluation Criterion (How to Judge) | Traditional Multi-Vendor Agency | In-House 17-FTE Team | 1 AI Systems Architect + Multi-Agent Pipeline (Our Model) |
|---|---|---|---|
| 24-Day Output Capacity | 150 to 300 hours (typically 1 discovery deck + 1 brochure site) | ~2,420 hours (after 3 to 6 months of recruiting and onboarding) | 2,420 verified billable hours across 8 live production properties |
| Equivalent Build Cost | $363,000 at $150/hr ($423,500 to $484,000 at $175 to $200/hr) | $145,000 to $220,000+ in monthly fully loaded payroll and benefits | $540.82 in compute & API spend + architect engagement fee or salary |
| Time to First Production Deploy | 4 to 8 weeks (account managers, wireframes, handoffs) | 8 to 16 weeks (recruiting, tooling, sprint ceremonies) | Same day (direct spec-to-code with automated rollback IDs) |
| AI Search (AEO) Readiness | Rarely included; heavy JS bundles, no .md mirrors or /llms.txt |
Requires dedicated SEO/AEO engineer to maintain manually | Built into CI/CD: 745+ .md mirrors, /llms-full.txt, Wikidata JSON-LD |
| Quality & Security Gates | Manual spot-checks; frequent regressions across handoffs | Depends on team discipline and test coverage | Deterministic: predeploy.sh, check-page-anatomy (11/11), 215 API tests |
| Account, Key & Code Ownership | Often locked in agency-owned hosting or proprietary templates | 100% company-owned | 100% client/company-owned git repos, Cloudflare, Twilio, and Stripe accounts |
The 8 live production web properties you can verify today
Every property in this 24-day audit is live on the public web right now. You do not have to take a slide deck on faith: you can curl the /llms.txt and .md mirrors, call the phone lines, and run the calculators yourself.
- Blur AI Solutions (bluraisolutions.com): 56 mirrored pages. Home of Nora, our 24/7 AI voice receptionist for home service contractors. Call the live demo line at (816) 683-2229 or inspect bluraisolutions.com/llms.txt.
- Blur Agentic (bluragentic.com): 17 mirrored pages + Cloudflare Pages Functions. Done-for-you speed-to-lead, estimate follow-up, and booking systems for roofing, restoration, and trade companies, starting with a $1,500 Growth Assessment credited in full toward any build. Inspect bluragentic.com/llms.txt.
- ContactLint (contactlint.com): 58 routes on a Cloudflare Worker backed by D1 SQLite and Durable Objects, with 215/215 passing Vitest assertions. Validates email syntax, MX/A DNS records, disposable domains, and phone numbers via REST API and a Streamable HTTP MCP server at
https://contactlint.com/mcp. Inspect contactlint.com/openapi.json and contactlint.com/llms.txt. - Kansas City SEO Services (kansascityseoservices.com): 74 mirrored pages. Google Map Pack, Google Business Profile, and AI answer visibility for Kansas City businesses, with published pricing ($0 Map Pack Check, $795 Listing Fix, $1,150/mo Map Pack, $1,850/mo Metro). Inspect kansascityseoservices.com/llms.txt.
- Email Marketing KC (emailmarketingkc.com): 72 mirrored pages including 21 technical and trade email guides. Done-for-you owner-voice email marketing for Kansas City businesses with existing customer lists. Inspect emailmarketingkc.com/llms.txt.
- KC Roofing Directory (kcroofingdirectory.com): 171 mirrored pages covering Kansas City metro cities, suburbs, ZIP codes, and roofing contractors checked against state records, paired with our Kate AI voice concierge. Inspect kcroofingdirectory.com/llms.txt.
- Crypto Guidance (cryptoguidance.org): 275 mirrored pages including 175+ plain-English cryptocurrency and self-custody definitions, tax and satoshi calculators, and our Bella voice concierge. Inspect cryptoguidance.org/llms.txt.
- Get Found with AI (getfoundwithainow.com) & johnalewine.com: Live multi-model AI recommendation testing across ChatGPT, Claude, Gemini, Grok, and Perplexity, plus 4 interactive Kansas City AI ROI and missed-call calculators at johnalewine.com/tools/.
Methodology: how we audited the 24-day output and compute ledger
To keep this case study grounded in verifiable engineering artifacts rather than marketing estimates, we followed five strict rules during the October 24, 2026 audit:
- Count only deployed production artifacts: Preview branches, discarded visual QA iterations, and temporary test directories were excluded. Only live routes, active voice agents, passing test suites, and production Worker/Pages deployments were counted.
- Conservative deliverable hour weights: Flagship custom pages and interactive tools were weighted at 6.0 hours, structured trade and guide pages at 1.5 to 3.0 hours, glossary definitions at 0.5 hours, custom Workers/MCP endpoints and test suites at 8.0 to 16.0 hours, voice agent telephony setups at 8.0 to 10.0 hours, and deterministic 60fps Canvas videos at 8.0 hours.
- Published agency hourly rates: We applied a $150/hr baseline for standard US software and digital agency work ($363,000 total) and compared it against the $175/hr to $200/hr specialist agency band reported by Haus Advisors / Promethean Research ($423,500 to $484,000).
- Reconciled token and billing ledgers: Model spend was pulled directly from
/workspace/logs/anthropic-spend.log(reconciled viallm_spend.pyagainst the Anthropic console calibration),/workspace/logs/vertex-spend.log, and/workspace/logs/xai-spend.log, plus fixed monthly subscriptions on Cloudflare, ElevenLabs, xAI Voice, and Twilio. - Machine-verified compliance: Every page in the portfolio was checked with
predeploy.sh siteand/usr/local/bin/check-page-anatomy.
Who this architecture is not a fit for
- Buyers looking for "$71 agency work": The $540.82 compute ledger ($71.64 net cash after model credits) is our internal infrastructure cost, not our retail service price. Our commercial engagements at Blur Agentic start with a paid $1,500 Growth Assessment (credited toward $5,000 to $10,000+ builds), and Nora at Blur AI Solutions is $897 setup + $297/month (or $0 setup on a 12-month plan).
- Teams that want unreviewed autonomous sends: In our architecture, external communications, pricing changes, and production promotions are gated by human approval and deterministic linters. If you want an unmonitored bot blasting unverified messages, we are not a fit.
- Foundation model pre-training research labs: We build applied multi-agent production systems, real-time voice telephony, MCP tool servers, and AI search architectures on top of frontier APIs rather than pre-training base weights from scratch.
- Anyone seeking guaranteed rankings or revenue promises: We guarantee deterministic engineering deliverables, go-live dates, and transparent math, never rankings or revenue outcomes.
Red flags when evaluating AI agent case studies
- No link to live production URLs: Screenshots of localhost dashboards or flowcharts with no public URLs, phone numbers, or API endpoints you can test yourself.
- No token or infrastructure ledger: Claims of "autonomous agent teams" that hide model selection, token burn, or voice-minute rounding.
- Zero deterministic verification gates: Relying on an LLM to "self-check" its own output without an external compiler, unit test suite (like Vitest), or DOM/schema validator (
check-page-anatomy). - Vendor lock-in on accounts and keys: Builders who refuse to deploy inside your own Cloudflare, AWS, Twilio, or Stripe accounts.
Frequently asked questions
How did one AI systems architect ship 2,420 billable hours in 24 days?
By separating human architecture from autonomous execution. John Alewine designed the system specs, data schemas, voice prompts, and deterministic verification gates (predeploy.sh, check-page-anatomy, and 215 automated tests), while specialized AI agents executed parallel tracks across web engineering, technical AEO, content research, voice telephony, data verification, and motion graphics.
How is the $363,000 agency output valuation calculated?
Every deliverable shipped between October 1 and October 24, 2026 was counted from git commits, deployed Cloudflare Pages and Workers, spend logs, and verified site trees, then multiplied by conservative human-equivalent deliverable hours (totaling 2,420 hours, or 17.3 full-time equivalents over 3.5 workweeks). At a standard $150/hr engineering and agency benchmark, 2,420 hours equals $363,000. At specialist US agency rates of $175 to $200/hr (Haus Advisors and Promethean Research 2026 benchmarks), the same output values at $423,500 to $484,000.
What is included in the $540.82 compute, API, and infrastructure ledger?
The $540.82 ledger covers every model token, voice minute, and serverless hosting bill from October 1 to October 24, 2026: $441.27 on Anthropic Claude API (covered by a $5,000 Anthropic API credit grant), $27.91 on Google Cloud Vertex AI (covered by a $300 trial credit), $22.00 on ElevenLabs Creator, $20.00 on xAI Grok Voice API, $12.87 on xAI Grok API search and evaluation calls, $11.77 on Twilio SIP trunking and local numbers, and $5.00 on Cloudflare Workers Paid. Net cash outlay after model credits was $71.64.
Does $540.82 in compute mean a client can buy agency or engineering services for $540?
No. $540.82 is the raw model token, telephony, and serverless compute ledger, not the price of a commercial engagement. Just as raw steel and electricity do not represent the price of a finished building, raw token spend requires a senior systems architect to design the workflows, write the verification harnesses, enforce security headers, and take production accountability.
How does the 19-Element Page Anatomy help sites get found in AI search (AEO)?
AI answer engines such as ChatGPT Search, Perplexity, Claude, Gemini, Grok, and Google AI Overviews prioritize pages that are cheap to parse, verifiable, and quotable. Every property in this build ships clean Markdown mirrors (.md), /llms.txt and /llms-full.txt corpora, explicit AI crawler allowlists in robots.txt, JSON-LD with Wikidata entity IDs, a 40 to 60 word quotable opening paragraph with sourced numbers, first-party HTML tables, and zero-login API endpoints.
How do you prevent AI hallucinations and regressions across 8 production sites?
Every build is gated by deterministic machine checks before deployment. Our predeploy.sh linter blocks banned phrases, punctuation violations, leaked credentials, and unapproved external calls. Next, check-page-anatomy verifies all 11 structural and AEO gates, and ContactLint runs 215 automated Vitest assertions before Wrangler deploys to Cloudflare Pages or Workers with a recorded rollback ID.
Can I hire John Alewine to build this architecture or join my engineering team?
Yes. John Alewine takes contract AI systems architecture and automation engagements through Blur Agentic and is open to full-time AI implementation engineer, forward-deployed engineer (FDE), and AI systems architect roles. Email [email protected] or call Nora live at (816) 683-2229.
Sources & verification endpoints
- Haus Advisors / Promethean Research: 2026 US Digital & Specialist Agency Hourly Rate Benchmarks ($150 to $200/hr)
- 1Digital Agency: 2026 Performance & Technical Agency Rate Benchmarks ($185/hr)
- o8 Agency: 2026 Blended Digital Engineering & Agency Pricing ($190/hr)
- Sentra: Anthropic Claude API Per-Token & Batch Pricing (October 2026)
- Imisofts: Twilio Programmable Voice & SIP Trunking Line-Item Pricing (October 2026)
- AppStack Insider: Cloudflare Workers & Pages Paid Plan Pricing (2026)
- johnalewine.com/llms.txt & johnalewine.com/llms-full.txt (Live AI Agent Corpus)
- ContactLint OpenAPI 3.1 Specification (contactlint.com/openapi.json) & bluragentic.com/llms.txt
What changed in this guide
- 2026-10-24: Published verified 24-day (October 1 to October 24, 2026) multi-agent engineering case study with 6-department hour table (2,420 hours / 17.3 FTE / $363,000 at $150/hr), $540.82 compute and API ledger, SVG architecture diagram, Wikidata JSON-LD entities, and clean Markdown mirror (/363k-agency-output-24-days-ai-agent-case-study/index.md).