Teardown No. 94 · Outbound

40+ Hours Off Cold Outreach: How Amrudin Catic split one prompt into four inspectable agents

Amrudin Catic built a cold outreach system for a client. It started as one prompt and ended as four sequential agents. The single prompt demoed well, but three runs on the same website gave three different conclusions. He split the work into four agents: one finds leads against the client's ICP, one audits the lead's website, one builds a personalized presentation, and one sends an email that names one specific audit finding. He says the system runs in production for a client and removes more than 40 hours of manual research and outreach work every week.

Amrudin CaticVerified Nov 7, 2026Markdown copy

The problem

The first version was one prompt that would crawl the website, analyze it and write the email. He says it “demoed beautifully.” Then he ran it three times on the same website: “Three runs, three different conclusions.” One pass fixated on technical SEO. The next talked about the design. A third ignored the most obvious problem on the page and wrote confidently about something trivial. In his words, “The failure mode wasn’t a bad email. It was variance.” Because the output was not reproducible, he could not evaluate it. He had no way to answer “when this sends a bad email, which part broke?”

The blueprint (as he describes it)

  1. Split the prompt into four agents. Each agent has one responsibility, produces structured output, and hands only that output forward. The pipeline is sequential. Each stage receives only the context relevant to its task. He puts the benefit this way: “Four agents give you three places to look.” When an email goes out wrong, he reads the audit output to see whether the observation was wrong or the writing was.
  2. Agent 1: find leads. It finds qualified leads against the client’s ICP. He writes, “A perfect email to the wrong company is still noise.”
  3. Agent 2: audit the website. It does not try to produce a 40-page SEO report. It looks across messaging clarity, value proposition, CTA visibility, UX friction, trust signals, branding consistency, content quality, basic technical SEO and conversion opportunities. Every finding passes two filters: “Is this actually true?” and “Would a business owner care?” Only findings that survive both move to the writing stage. He says most do not.
  4. Agent 3: build the presentation. It generates the personalized presentation from the outputs of agents 1 and 2.
  5. Agent 4: send one specific sentence. It sends a conversation-starter email that references one specific finding from the audit by name. He says “Every other component exists to make one sentence in one email specific and credible.”
  6. Give the writing agent less. He first tried better prompts, with more instructions, more examples and more careful phrasing about tone. “It never worked.” The fix was to reduce the model’s freedom. The writing agent never sees the whole website. It receives one or two validated observations and a strict set of writing constraints. It is not allowed to exaggerate, to invent compliments, or to stack multiple insights into one paragraph.
  7. Fall back instead of inventing. If the audit cannot find a meaningful observation with enough confidence, the system falls back to broader business context and sends a less personalized email. It does not make up a problem.
  8. Spend the effort on the input. The bottleneck was websites, not prompts. He lists heavy JavaScript rendering, inconsistent HTML, cookie banners, page builders, content duplicated across templates and missing metadata. As he puts it, “Feed that to a good model and you get a confident audit of a cookie banner.” Most of the engineering went into context: extracting real content from rendered pages, normalizing inconsistent markup and deciding what to throw away. “The AI part took days. The context part took months.”
  9. Keep the stack boring. He uses Next.js with TypeScript on Vercel, with an orchestration layer that coordinates the agents in sequence. He uses OpenAI models and picks a different model for each task to balance quality, speed and cost. “Most steps don’t need the smartest available model. They need a predictable one.”

Stated results

  • In production for a client, the system “removes more than 40 hours of manual research and outreach work every week.”
  • He says restricting the writing agent to one or two validated observations produced more human-sounding emails than any prompt refinement did. No measurement is given.
  • Reply rates, meetings booked, revenue and email volume: None stated.

Caveats

  • This is a first-person post on his own site, and he sells this kind of work. The post says “that’s the work I do” and offers a 20-min call. It also points readers to an SEO scorer and lead scorecard on his site.
  • The 40+ hours figure is self-reported for an unnamed client. The source does not say how it was measured. This is client work he built, not his own outbound.
  • The full engagement is written up in an AI agents case study that is not part of the source text.
  • The AI layer uses OpenAI models, but the specific models are not named. The source does not name the lead data source for Agent 1 or the tool that sends email.
  • If he rebuilt it, he says the time would go into structured data collection, better website parsing, richer business context and feedback loops from actual outreach performance.
  • Handles: LinkedIn /in/amrudincatic (verified 2026-11-07: linked from amrudincatic.com footer and llms.txt). GitHub amrudincatic (verified 2026-11-07: site and llms.txt). Instagram @amrudincatic (verified 2026-11-07: site footer). X, Bluesky, Threads, TikTok, Mastodon: none linked from primary surfaces.

FAQ

Why did he use four agents instead of one prompt?

One prompt gave three different conclusions across three runs on the same website. Four agents with structured handoffs let him trace a bad email to the stage that caused it. He calls the benefit attribution, not intelligence.

What does the writing agent actually see?

It never sees the whole website. It gets one or two validated observations and strict constraints. It may not exaggerate, invent compliments or stack multiple insights into one paragraph.

What happens when the audit finds nothing useful?

The system falls back to broader business context and sends a less personalized email. It does not make up a problem.

Does he report reply rates or meetings?

No. The only stated result is that the system removes more than 40 hours of manual research and outreach work every week for a client.

Sources

Want agents like these working your distribution? Get It Built. Get new teardowns weekly: subscribe (double opt-in).

Get one teardown a week.

Same format, a new operator every issue. Double opt-in.