Autonomous agent chains that build, test, and improve themselves. No hardcoded vendors. No human in the loop. Just typed contracts, tournament selection, and relentless self-optimization.
Agents that just generate text are useless. Real agents need to ACT: scrape, send, track, evaluate. Each step needs the best possible approach -- not hardcoded vendors.
Today's best email API might be tomorrow's worst. Agents wired to a single provider can't adapt when the landscape shifts. The Foundry treats every tool as a replaceable slot.
Most agent pipelines are fire-and-forget. Without evaluation loops, you never know what's working. The Foundry runs Karpathy loops on every slot to surface winners.
Humans shouldn't wire APIs together by hand. The Foundry composes chains automatically from tournament-winning micro-agents, then executes them slot-by-slot.
Five stages transform a business intent into a self-optimizing autonomous workflow. No code. No vendor lock-in. Just contracts and competition.
Each slot declares a typed contract: what goes in, what comes out, and which KPIs measure success. Think of it as an interface that any micro-agent can implement. Input schema, output schema, evaluation criteria -- all explicit.
The system scans the AaaS Vault -- 4,230 skills across 52 categories -- to find every possible approach for each slot. Firecrawl vs. Apify for scraping. Resend vs. SendGrid for email. All candidates surface automatically.
Each candidate micro-agent competes. The Karpathy evaluation loop runs them against real data, scores results on the slot's KPIs, and assigns Elo ratings. The best approach rises to the top. Losers get archived.
Tournament winners from each slot are wired together into complete chains. The email chain gets the best lead finder, the best enricher, the best writer, the best sender -- each independently proven.
Chains run slot-by-slot, fully autonomous. Each slot receives the previous slot's typed output, processes it, and passes structured results downstream. Monitoring and metrics are built into every execution.
Current production metrics from the Micro-Agent Foundry.
Each chain family is a complete business workflow decomposed into typed slots. Tournament winners fill every position.
The most complete chain: 6 slots, each filled by a tournament-proven micro-agent. Here is the full pipeline with the real APIs powering each slot.
Discover target companies and contacts from web scraping and search APIs
Crawl company domains to extract context, tech stack, and decision-maker info
Generate personalized multi-step email sequences from enriched lead context
Deliver emails with tracking, unsubscribe headers, and rate limiting
Track opens, clicks, replies, bounces, and unsubscribes per sequence
Karpathy loop scores sequence performance and triggers re-tournament if below threshold
Three ways to activate the Foundry. Each triggers a full chain execution with tournament-proven micro-agents.