Chris — the outbound AI agent you asked for. I shipped v1.0 this morning.
You asked: "How do I build an AI agent that suggests warm leads + recommends email/LinkedIn/phone activity, without mass-blasting?" Universal timing — I literally shipped v1.0 of this exact thing today. This page is the architecture, the tool list, the doctrine, and the two paths to get it running for your prospects. No Apollo subscription. No Instantly. Free to run if you build it yourself, $100 if you want me to deploy it on your machine in a working session.
The agent doesn't auto-send. It pulls signal from where buyers actually live (LinkedIn engagement, news mentions, technical Q&A demand), ranks who needs a touch right now, drafts the email/DM/text in your voice, and routes every draft through you for one-click approve. Free to run on your machine (Python + free APIs + a Cloudflare Worker for the send rail) — no SaaS subscription, no per-seat cost. Built it because every "AI SDR" I tried was a mass-blast tool wearing a personalization mask. This is the opposite: tiny volume, near-100% land rate.
1What the agent actually does
That's exactly what it does — plus one thing you didn't ask for that's load-bearing: it drafts the actual outbound text for each lead, in your voice, and never sends without you clicking approve. The drafting part is what separates this from "give me a CSV of leads." The CSV is a starting point. The draft is the work.
- Pulls signal: LinkedIn engagement (who liked/commented your last post + your competitor's posts), news mentions of target companies (funding, hiring, awards, leadership changes), technical Q&A demand (Reddit + StackExchange threads in your vertical).
- Ranks who needs a touch right now: scores each lead on signal recency + email readiness + cluster fit + GSC/SEO-page match (if relevant to your business).
- Recommends channel: LinkedIn DM if they engaged on LinkedIn this week, email if you have a verified email + a real signal (news), phone for the top-3 ranked accounts only.
- Drafts in your voice: uses your prior outbound (LinkedIn DMs, sent emails) as voice training, generates a 2-3 sentence draft per lead, references the actual signal that triggered the touch.
- Never auto-sends: every draft routes through your dashboard. One-click approve, one-click skip, one-click snooze 7 days.
2The architecture · 5-layer flow
Pull from where buyers live
RSS feeds, news watchers, LinkedIn engagement scrape (compliant rate-limit), Reddit+StackExchange tag watches. Outputs structured JSON: {name, company, signal_type, signal_text, recency}.
Rank who needs a touch now
Each lead gets a score: signal recency (+50 if <7d), email-ready (+20), status not-recently-touched (+12), cluster fit (+3 per matched keyword). Top-N surface in the dashboard.
Pick the right surface
LinkedIn DM if last engagement was LinkedIn-side. Email if you have a verified email + signal worth referencing. Phone reserved for top-3 only — your time is the constraint, not their preference.
Write in your voice
LLM call (Claude Sonnet or GPT, whichever you prefer) with your prior outbound as voice samples + the lead's signal as context. 2-3 sentence draft. Hard-codes your sig + phone.
Human-in-the-loop, always
Dashboard shows each draft. One-click approve fires through your send rail (email Worker, LinkedIn copy-paste, phone-ready dial). One-click skip kills it. One-click snooze 7d sleeps the lead.
3The tool list — what it actually runs on
feedparser for RSS, requests for news APIs, lightweight LinkedIn engagement scraping (manual export or API)4Honest 80/20 — where this works and where it doesn't
It's a precision tool, not a volume tool
This agent is built for the operator who wants 5-20 thoughtful touches a day, near-100% land rate, no risk to their domain reputation, and no time spent inside a SaaS dashboard. Where it doesn't work:
- If you need 500-1,000 sends a day, you need a real outbound platform with deliverability infra, IP warming, and bounce handling. This is the wrong tool.
- If you don't have a clear ICP yet, the agent can rank but it can't define who matters. The ICP is a 30-min upfront conversation.
- If you need a multi-rep team to run it, the dashboard isn't multi-tenant yet. v1.0 is single-operator. v2 could expand if there's pull.
Where it works: founders, owner-operators, fractional consultants, recruiters, anyone whose deal value is high enough that 5-20 thoughtful touches/day beats 500 mediocre ones. KD Sweetwater (talent systems consulting) sounds like exactly that shape.
5Two paths to running it for your prospects
Architecture write-up + tool list
I send you the full architecture doc, the tool list with API setup steps, the Python scaffolds for each layer, and the LLM prompt templates I use for drafting. You build it on your machine over a weekend.
I deploy it on your machine in a working session
60-90 min screenshare working session on your machine. I deploy the v1.0 stack against your actual prospect base, set up the news watcher with your ICP keywords, configure the LLM voice training from your prior outbound, and you start using it tomorrow.
Either path, async, no sales call
Reply on LinkedIn or text me. If DIY, I send the doc + scaffold within an hour. If Quick Clarity, we pick a 60-90 min slot this week, you have the agent running by end of session.