Built an autonomous prospecting engine that fuses a research swarm, a writing agent grounded on the prospect company, and a deliverability supervisor. Replaces the work of a 4-SDR team on a per-account basis.
The Challenge
This early-stage outbound sales agency needed to scale personalized B2B outreach across forty distinct client accounts simultaneously without hiring a massive SDR team. Templated email marketing and LinkedIn outreach sequences generated extremely low reply rates of under 2% and triggered high domain blacklisting risks due to over-sending. The agency required an autonomous system that could research prospect organizations, extract technographic data, write highly personalized messaging that read as human-written copy, and rotate senders safely. The primary technical challenge involved building a multi-tenant orchestration pipeline that could store, analyze, and optimize campaign learnings. It was critical to maintain strict database isolation across client vectors and tenant spaces to prevent sensitive target metrics and proprietary campaign data from leaking between competing clients.
Pain points we set out to solve
- ×Sub-2% reply rates on templated cold sequences
- ×High domain blacklist risk from over-sending
- ×Manual research took 20+ minutes per account
- ×No feedback loop from replies back to copy improvement
Objectives
- 01Push qualified reply rates above 8% on seeded cohorts
- 02Research an account end-to-end in under 60 seconds
- 03Keep per-domain send rates under deliverability thresholds
- 04Let each client tenant keep its data and learnings isolated
Approach
How we delivered — phased, with clear checkpoints and evidence at each step.
- Week 1-3
Agent graph design
Mapped the outbound workflow into four agent roles - researcher, writer, scheduler, and supervisor - wired through a LangGraph state machine with explicit handoff contracts and retry logic.
- Week 4-7
Research swarm
Built a parallel research swarm that scrapes the prospect site, pulls LinkedIn via Playwright, summarizes recent news, and detects the tech stack via Wappalyzer. Outputs a structured brief the writer uses as grounded context.
- Week 8-11
Writer and RAG corpus
The writer agent drafts email and LinkedIn copy grounded on the research brief plus a per-tenant RAG corpus of prior winning sends. No generic templates - every send is composed from first principles for the account.
- Week 12-14
Deliverability supervisor and feedback loop
A supervisor checks every draft for spam triggers, enforces per-domain daily caps, and rotates sender identities. A reply-ingestion job tags responses by outcome and feeds them back into the RAG corpus as positive or negative examples.
The Solution
Four agents on a LangGraph state machine handle the full outbound loop: research -> write -> schedule -> supervise. Each tenant has its own RAG corpus of prior wins, so the writer agent compounds on what works for that client without cross-contamination. The supervisor is the deliverability brake that keeps campaigns safe. The system rotations and daily schedules ensure compliance with major email hosts.
Research swarm
Parallel agents pull site content, LinkedIn profile, recent news, and tech stack into a structured brief in under 60 seconds per account.
RAG-grounded writer
Every send is composed from the research brief plus retrieved examples of prior wins from the same tenant corpus - no generic templates.
Deliverability supervisor
Validates spam triggers, enforces per-domain daily caps, rotates sender identities, and pauses campaigns automatically when reply quality drops.
Multi-tenant isolation
Each client gets its own Postgres schema, vector namespace, and sender pool - so learnings and data never leak between tenants.
Reply-outcome feedback loop
A daily job tags inbound replies as positive, neutral, or negative and mints training examples for the RAG corpus, so the system gets sharper every week.
Technology stack
Picked for latency, cost, and long-term maintainability — not for novelty.
AI / Agent
Scraping / Research
Messaging / Infra
Observability
Results
Business impact
The agency onboarded 12 new clients in the quarter without adding headcount. Two existing clients renewed at 2.5x prior spend after seeing reply-rate lift in the first 30 days.
Key takeaways
- Outbound writing quality is bounded by research quality - invest in the research swarm first
- A deliverability supervisor is non-negotiable at scale, otherwise you will lose a domain before you learn what did not work
- Multi-tenant RAG corpora turn each client into a compounding advantage instead of a one-shot campaign
Visual reference
