The high-leverage teams

About a hundred engineering teams that get outsized output, grouped by how they do it — the same nine leverage patterns we found in individuals, applied to organisations. Then: what running our own research toolkit to build this taught us about the toolkit itself.

This is a curated sample, not a ranking. There is no objective "top 100 teams," so the value is the grouping, not the membership or any order. Assembled with our research system (GitHub and news signals) plus judgment; team placements are by their primary lever, and many pull several.

Open-source compounding

Effort multiplied by everyone who runs your code.

• Linux Foundation — Linux and Git, the substrate under everything
• PostgreSQL, SQLite, ClickHouse — the databases others build on
• curl (Daniel Stenberg) and FFmpeg — in billions of devices
• Astral — Ruff and uv; the fast Python toolchain
• Oven — Bun; runtime, bundler, package manager, test runner in one
• Zig Software Foundation and Ghostty — systems craft in the open
• Redis, Home Assistant, Blender, Godot — category-defining projects
• Python core, Rust Foundation, pandas — the languages and tools of the field

Framework and abstraction

Effort multiplied by every app built on your abstraction.

• 37signals — Rails, and the one-person framework
• Vercel — Next.js and the frontend-cloud pattern
• Laravel — the PHP framework plus a paid ecosystem
• Supabase — the open Firebase alternative on Postgres
• VoidZero (Evan You) — Vite and the JS build layer
• Svelte, Deno, Prisma, Remix — framework-shaped leverage

Agent-native development

Effort multiplied by parallel AI agents. The ascendant pattern — and, on GitHub today, the one dominating the trending repos.

• Anthropic — Claude Code, and "antfooding" it internally
• Anysphere — Cursor
• Cognition — Devin
• GitHub — Copilot
• Sourcegraph — Amp and code search
• Replit, Windsurf, Poolside, Augment, Factory — agentic coding
• Zed Industries and Warp — the editor and terminal, rebuilt for agents
• Aider and Cline — open-source coding agents

Taste and the small elite team

Output multiplied by a high bar and no coordination drag.

• Linear — small senior teams, no handoffs, taste over metrics
• Figma — deep technical craft behind a simple surface
• Raycast, The Browser Company, Superhuman — opinionated, fast
• Notion, Height, Campsite, 1Password, Craft — quality as the moat

Solo and tiny-team automation

One or two people running what used to need a company.

• levels.io (Pieter Levels) — dozens of profitable products, nearly alone
• Marc Lou — a portfolio of small SaaS
• Plausible, Bear, Pinboard — tiny teams, durable products

Community and distribution as leverage

Effort multiplied by reach — an audience that adopts, extends and spreads what you ship.

• Hugging Face — the hub the open-model world runs on
• Ollama — local models for everyone
• LangChain, CrewAI — agent frameworks with huge communities
• n8n, Dify, ComfyUI, Firecrawl — community-driven AI tooling topping GitHub

Radical simplification and boring tech

Output multiplied by removing what you'd otherwise maintain.

• Fly.io, Railway, Render — deploys without a platform team
• Tailscale — networking that just works
• Turso / Litestream (Ben Johnson) — SQLite everywhere
• PlanetScale, Oxide Computer — hard problems, few moving parts

Capital and compute: the frontier labs

Leverage via money, chips and concentrated talent.

• OpenAI, Anthropic, Google DeepMind — the leading labs
• Meta AI (FAIR / Superintelligence Labs), xAI — big-company bets
• Mistral, DeepSeek, Alibaba Qwen — open-weight frontier
• Safe Superintelligence, Thinking Machines — the well-funded stealth labs
• Nous Research, Together AI, Fireworks, Unsloth, vLLM, Cohere — the open and infra layer

Own-the-stack and hardware

Leverage by controlling every layer, so nothing external caps you.

• SpaceX, Tesla (Autopilot, Optimus) — full-stack engineering at scale
• Nvidia, Apple Silicon — the hardware the field runs on
• Cloudflare, Modal — own the edge and the compute
• Anduril, Figure, comma.ai, Boston Dynamics — software-defined hardware
• Isomorphic Labs — DeepMind's drug-discovery spinout

Lessons from running the research system

This list, and the research pieces before it, were assembled by running our own collection toolkit many times in one day. That was itself the most useful test of it. What the runs taught us, refined:

• Collection is nearly free; judgment isn't. Structured public sources hand over clean metadata, so gathering costs no model tokens, and local dedupe plus ranking cut what a model has to read by about 97%. But no amount of collection produces a ranking — every "top N" here is judgment, and pretending otherwise would be the dishonest part.
• Meta-search is broad, not bottomless. Our SearXNG hit its upstream engines' rate limits twice under a day of heavy use — Brave, DuckDuckGo and Google all throttled. A serious research cadence needs structured APIs as the backbone and a paid high-quota search as backup, not a free meta-search alone.
• GitHub stars are a live, cheap signal. Today's trending repos are dominated by agent tooling — Claude Code, OpenCode, Firecrawl, Dify, ComfyUI. The agent-native wave isn't a vibe; it shows up in the data we pulled.
• Counts mislead on "who builds X." Keyword and citation search surfaced medical-LLM papers over the people actually building frontier models. For "who's doing the important work," a count is a trap; human judgment on public output is the only honest ranker.
• Access is a ladder, and the best sources are often walled. The highest-signal forums and platforms sit behind bot walls. The honest move is to record the wall with a date and route to a sanctioned API or feed, not to defeat it. Our policy held all day: api where there's an API, feed where one is published, respect otherwise.
• Agent-written output needs the grounding check. The report writer invented a citation on its first run; our checker caught it because the link matched nothing we'd collected. Verification is not optional once an agent is in the loop.
• The honest limit beats the impressive claim. Across every run, "here is what we could reach, and the policy and evidence behind it" was worth more than "we got everything." That's the difference between a scraper and a research practice — and it's the one lesson that didn't change all day.
• Reanalysis (our own tradecraft, applied back). Two blind spots we now name: this list mirror-images toward English-visible, Western teams, so it undercounts Chinese labs (DeepSeek, Qwen) and work done quietly inside large companies; and the GitHub-stars signal that flagged the agent-native wave is gameable, so read it as direction, not magnitude.

Provided as is, for information only, with no warranty and no liability. Not legal or professional advice. Security and radio material is for authorized, lawful use only. Full disclaimer.

Visits are logged on our own server: IP address, browser and device details, and pages viewed. We use them to understand traffic and spot attacks, and never share or sell them.