Founder / 2025 to 2026 / Wound down June 2026

Magnetite

Multi-agent outbound automation. Four agents source, grade, build and deliver a real piece of work for every prospect.

TypeScript
Multi-agent pipelines
MCP (~50 tools)
Thompson sampling
OpenClaw
Remotion
Retrieval-grounded generation

The output. A personalised blueprint for one prospect, fronted by a custom video sales letter, with the claim and booking flow built into the page.

01 · Problem

B2B cold outreach gets reply rates under 2%. What actually works is giving real service value before asking for anything, and doing that properly costs an hour or more of skilled human time per prospect. So it either does not scale, or it degrades into spam.

Magnetite did the whole prospecting job instead of assisting a human with it. Not competing with research tools like Clay or Apollo. Replacing the pitch itself.

02 · What I built

Four cooperating agents. One sourced companies. One graded whether the company and the specific person were the right target, before any expensive work happened. One produced a real piece of work for the prospect, a genuine deliverable rather than a compliment. One packaged and delivered it.
Deliverables were the kind of work a service business normally charges for. Custom audits, competitive analysis, ROI calculations. Personalised to each prospect and produced in minutes, not hours.
An API plus Zapier and n8n integrations, so the lead magnets plugged into whatever outreach channels the customer already ran.
A Thompson sampling bandit decided source spend. Budget flowed automatically to whichever sources were producing accepted targets.
Retrieval-grounded generation. Outputs were written from material the research step actually gathered, never from the model's imagination.
An MCP tool layer of around 50 tools that every agent worked through.
Per-customer isolated OpenClaw instances, each with its own credentials, data and browser sessions. Nothing shared between tenants.
A Remotion video pipeline for custom video deliverables.
Cost per delivered unit of work tracked as a first-class metric, with the grading agent acting as a cheap gate in front of expensive downstream work.

03 · Architecture

The pipeline is a budget-guarded funnel. Cheap decisions happen early, expensive work happens late, and every stage can reject. A bandit tunes where the money enters.

04 · What happened

The platform went from nothing to working product in three months, solo. Real customers were onboarded and supported by me personally, and one customer measured a 17.67% reply rate across 20,000 emails, against an industry standard under 2%. That number is the whole thesis. Give real value first and cold outreach stops being cold.

Not everything held. An early iteration used a custom LinkedIn scraper with residential IP rotation and cookie transfer. It worked, and it did not keep working for longer than about a month. That failure drove the move to per-customer isolated browser sessions.

Magnetite also did not end where it started. In 2026 it evolved into the Proof Engine, which is what magnetite.ai shows today. The insight: your customers already wrote your best ads. The engine swept LinkedIn, X, Reddit, YouTube, podcasts and review sites daily for genuine praise, including the untagged mentions social listening tools cannot see, graded every unit by what actually converts, cleared usage rights with each author, and ran the best of it as LinkedIn Thought Leader Ads. Consent was architecture, not policy. An ad could not render unless the rights state was cleared, enforced in code.

I wound Magnetite down in June 2026. What survives is the way I now build everything: grading gates in front of expensive work, per-tenant isolation, consent enforced in code, and cost per unit of work as a number you watch daily.

In the product

Magnetite's create project screen, where a customer describes their business, audience and value proposition, and defines the API input schema the research agent works from. — A project starts as a brief. Audience, problem, value proposition, and the input schema the research agent will be fed for each lead.

The research agent builder showing a planner and executor step, tool assignments like web search and website tech scanner, a mandatory loop evaluator, and a per-execution credit cost breakdown. — The research agent. A planner and executor loop with real tools, a mandatory evaluator deciding whether to loop again, and the cost of every execution priced up front.

The Magnetite Prompt Studio, showing the generation prompt, available variables, temperature settings and the video sales letter pipeline. — The Prompt Studio. The generation prompt, its variables, model settings and the video sales letter pipeline behind every deliverable.