Autonomous AI agents that discover, qualify, and close B2B deals
Specialized LangGraph agents run your sales motion — discover companies from job postings, enrich profiles with grounded DeepSeek reasoning, find decision-maker contacts, and draft personalized outreach. Every send waits for your approval. Cron-driven, traced, and eval-gated.
- 6
- touch cadence 4/7/7/7/7-day gaps, stop-on-reply
- 100%
- human-approved sends draft-first — nothing sends without you
- 80%
- eval accuracy gate deepeval bar on every prompt/model change
- 24x7
- autonomous agents cron-driven langgraph threads
Seven autonomous stages, one coherent pipeline
from job postings to human-approved outreach — on DeepSeek + Cloudflare D1.
Orchestration
LangGraph Python agents, checkpointed in Cloudflare D1 — resumable, durable, one thread per campaign + contact
LangGraph Python agents, checkpointed in Cloudflare D1 — resumable, durable, one thread per campaign + contact
Discovery
company_discovery agent sources companies from job postings — grounded, no synthetic leads
company_discovery agent sources companies from job postings — grounded, no synthetic leads
Enrichment
company_enrichment + problems agents reason over each profile with DeepSeek V4-Pro via the Cloudflare AI Gateway
company_enrichment + problems agents reason over each profile with DeepSeek V4-Pro via the Cloudflare AI Gateway
Contacts
contact_discovery + find_decision_maker agents identify and verify decision-maker emails
contact_discovery + find_decision_maker agents identify and verify decision-maker emails
Scoring + QA
opportunity classifier + company_qa audit score and gate every lead against grounding checks
opportunity classifier + company_qa audit score and gate every lead against grounding checks
Outreach
durable campaign engine drafts a 6-touch sequence — draft-first, sent only on your approval via Cloudflare Email
durable campaign engine drafts a 6-touch sequence — draft-first, sent only on your approval via Cloudflare Email
Evaluation
DeepEval ≥ 80% accuracy gate on every prompt/model change, with full LangSmith tracing across agents
DeepEval ≥ 80% accuracy gate on every prompt/model change, with full LangSmith tracing across agents
The real stack, no fabricated benchmarks
No local GPU farm, no invented numbers — just LangGraph agents, DeepSeek reasoning, and Cloudflare D1, eval-gated and traced.
eval accuracy gate
DeepEval bar enforced on every prompt/model change
human-approved sends
draft-first — nothing sends without your approval
touch cadence
4/7/7/7/7-day gaps · stop-on-reply
single source of truth
cloudflare d1 (sqlite) · drizzle orm
qdrant retrieval
agentic_rag · fastembed bge-small
langsmith tracing
every agent run traced end-to-end
typed api
apollo server 5 · codegen contracts
open source
fork it, self-host it, extend the agents
Three systems behind the autonomous pipeline
Autonomous LangGraph agents
[ checkpointed & resumable ]
Discovery, enrichment, contact-finding, and outreach run as Python LangGraph agents — checkpointed in Cloudflare D1, so every thread is durable, resumable, and cron-driven. No manual babysitting.
- —~36 specialized graphs registered in the backend
- —interrupt-and-resume keeps long sequences alive across days
Grounded, eval-gated AI
[ ≥ 80% accuracy gate ]
DeepSeek V4-Pro reasoning routed through the Cloudflare AI Gateway, with schema-constrained outputs and grounding checks. Every prompt or model change is tested against a DeepEval accuracy bar before it ships.
- —schema-constrained outputs, validated against the taxonomy
- —full LangSmith tracing across every agent run
Human-approved outreach
[ draft-first, always ]
A durable 6-touch campaign engine drafts every email, then waits. Nothing sends without your approval, the sequence stops the moment a reply lands, and delivery runs through Cloudflare Email.
- —approve, edit, skip, or reject each touch before it sends
- —suppression, caps, and cadence guardrails on every campaign
Read the architecture in CLAUDE.md and the backend graph registry.
browse the sourceAgentic Sales is fully open source — fork it, self-host it, extend the agents for your ICP.
Stop managing pipelines. Let agents do it.
Deploy once, run forever. Your agents discover, enrich, score, and draft outreach around the clock — you approve every send.
▍ not ready yet? get pipeline updates