Not a chatbot.
An agent that ships outreach.
An autonomous system that researches prospects, enriches them, drafts personalized outreach in your voice, queues for your approval, sends across channels, and routes replies — all on a schedule, all logged, all extensible. Built like production software with retries, observability, and an explicit human-in-the-loop. Demos don't survive contact with real volume. This is built to.
From ICP seed to reply in your CRM.
Below is the agent architecture I'd reach for on a production outreach system. Specialized agents do bounded work, an orchestrator decides what runs when, a queue handles state, and the human approval step sits exactly where a person actually adds value — not on every step. Hover any node to see what it solves and how it works.
Five agents. One production system.
Each agent ships with a defined input contract, output contract, tool whitelist, and observability. The system isn't done when it generates one good email — it's done when it runs unattended for two weeks and the results are still good. That's the bar I'm building to.
- ICP definition workshop — codify your actual fit criteria (not just industry + headcount)
- Voice library — past emails, sales calls, sample replies captured as a reusable style guide
- Channel inventory — where you currently send, deliverability state, daily volume caps per channel
- CRM audit — current shape, custom fields, what the agent writes back where
- Research agent — web search, site scrape, LinkedIn signals, recent news, role-specific hooks
- Enrichment agent — Apollo for verified contacts, Clay for firmographics, custom APIs where useful
- Bounded tool access per agent — no agent gets a tool it doesn't need
- Output contracts — structured JSON every downstream step can rely on
- Qualify agent — score 1–10 against your ICP with explicit reasoning attached, disqualify low scores
- Draft agent — personalized email + LinkedIn variants, in your voice, using the research as concrete hooks
- Self-critique step — the agent reviews its own draft against your style guide before queueing
- Anti-template safeguards — flag drafts that look like template spam before they reach approval
- Approval UI — approve, edit, skip; bulk-approve for high-confidence drafts
- Channel routing — email via Instantly/Smartlead, LinkedIn via your preferred path, drip cadences
- Sender reputation guardrails — per-channel daily caps, warmup respect, deliverability monitoring
- Send logs everywhere — every action attributable to a prompt version, a draft, an approver
- Reply classification agent — interested, objection, not now, unsubscribe, OOO, wrong person, spam
- Auto-actions per class — stop sequence on unsubscribe, surface "interested" to your inbox immediately, log OOO and retry later
- CRM sync — every prospect, draft, send, reply, and outcome written to your CRM with full attribution
- Scheduled runs — research nightly, draft daily, send during business hours, all on its own clock
- Observability dashboard — daily metrics, agent health, prompt costs, reply rates, sender rep per channel
- Runbooks + Loom walkthroughs — your team can intervene, adjust ICP, swap channels, ramp volume without me
Five weeks to live. Then we keep going.
Working software ships to staging every Friday. By week three you're approving real drafts on real prospects. By week five the system runs unattended and you're reviewing outcomes, not nursing pipelines. Then we expand.
Week 1 · Discovery + Foundation
ICP, voice library, channel + CRM audit, orchestrator scaffolded, infrastructure baseline live (queues, secrets, observability).
Aligned + scaffolded
Week 2 · Research + Enrich
Research agent finding prospects against your ICP. Enrichment wired. First end-to-end run on 10 real prospects by Friday.
Pipeline filling
Week 3 · Qualify + Draft
Qualify agent scoring with reasoning. Draft agent writing in your voice. You're approving real drafts on real prospects by Friday.
Drafts in queue
Week 4 · Approval + Multi-Channel Send
Approval UI live, sends going across channels with deliverability guardrails. First real campaign live by end of week.
Sending live
Week 5 · Reply Handling + CRM Sync + Schedule
Reply classification, CRM write-back, scheduled autonomous runs. System running unattended. Handoff doc + observability dashboard delivered.
Autonomous
Ongoing · Iteration + New Agents
Prompt tuning from real reply data. New channels. New ICPs. New agents (e.g., meeting prep, follow-up nurture, account-based plays). Partnership mode.
Partner mode
Pick 3 real prospects. I'll show you the agent's output.
A 30-minute call where you send me three prospects from your ICP — names + companies. I'll walk through how the research agent would approach them, what the draft agent would produce, and where you'd want the human approval gate to sit for your team. Live, on call. Not a slide.