Hiring an SDR is expensive. Salary, software seats, onboarding, ramp time — a single rep in their first year easily clears $80k before they've booked their first qualified meeting. And the brutal part: most of that cost isn't in the selling. It's in the research, the personalization, the sequencing — the mechanical work that happens before any real conversation starts.
That's the part that's now automatable. Not the relationship. Not the closing. The top-of-funnel research-and-outreach loop that burns most of an SDR's actual week.
The founders who've figured this out aren't hiring SDRs for prospecting work anymore. They're running what I'd call a Ghost Employee workflow — a multi-agent system that handles the mechanical parts of outbound, from prospect research through first send, without a human doing any of it.
Here's how it's actually architected.
#The three-stage structure
The workflow has three stages, each handled by a specialized agent with a specific job. They pass context forward to the next stage. They don't try to do each other's work.
Stage 1: Research. The first agent starts with a company name and domain. It queries Apollo or Crunchbase for firmographic data, scrapes recent blog posts and press releases, checks executive LinkedIn activity, and flags trigger events — recent funding, a new product launch, a leadership hire. The output is a structured JSON object: what this company cares about right now, what's changed recently, what problems they're likely sitting on.
This stage is entirely deterministic. The agent calls tools that return real data. No LLM involved yet — and that's intentional. You don't want a model hallucinating company facts that end up in the email.
Stage 2: Strategy. Now the LLM enters. The strategist agent takes the research output and matches the prospect's current situation against your value proposition. It doesn't write anything yet — it decides the angle. Something like: "They just raised a Series A and are building out engineering fast. Our dev-tool reduces new hire onboarding time. Lead with that."
The output is a strategic directive, not a draft. This separation matters more than it sounds. A lot of people collapse stages 2 and 3 — asking the model to research-and-write in one prompt. The output is generic. Separating research from angle from writing produces emails that feel like someone actually looked at the company before reaching out.
Stage 3: Writing and send. The writer agent takes the directive and produces a 3-sentence cold email. No templates. No "Hope this finds you well." The email references something specific — the Series A, the recent blog post, the engineering hiring spree — because stage 1 actually found it.
The email queues in your sending tool (Instantly or Smartlead) but doesn't fire automatically. Not at first. For the first few weeks, read 20-30 of these before you trust the system enough to remove that step. The failure modes are real and you need to see them while you can still catch them cheaply.
#What makes this work and what breaks it
The reason agent-driven outreach actually converts — when it does — is synthesized context. The prospect can tell someone looked at their company before reaching out, not just found their name in a list. Generic AI outreach with variable insertion is already dead. Buyers are immune to it.
What breaks the system: bad data sources, stale APIs, prospects with sparse online presence, and over-relying on the writing agent to compensate when the research comes back thin. When stage 1 doesn't find real, recent, specific information, the writing agent shouldn't try to fill the gap with plausible-sounding assumptions. It should escalate to a human.
Build that escalation logic in from the start. Any run where the research output is below a confidence threshold should route to a human review queue, not straight to the writer.
#Building it without an engineering team
The infrastructure is Make.com or n8n for orchestration, with OpenAI or Anthropic API nodes handling the LLM stages. Research tools — Firecrawl, Apollo, Crunchbase — connect as HTTP modules. Results write to an Airtable base or Notion database for monitoring before anything sends.
Start with 5 prospects, manually review every output, and find where the system breaks. Then 20. Then 100. Don't skip the observation phase — the failure modes surface early, and they're fixable when you're watching closely enough to catch them.
Stop scaling headcount to solve a problem that's now a system-design problem.