Experiments
-
An AI Agent Designed a 2 MAC per Clock INT4 Accelerator for One Tiny Tapeout Tile, Without Spending Money
Over two days the lab used an AI agent pipeline (chip lab, driven by Claude Code) to take a small dot product accelerator from a UART toy to a weight stationary INT4 design that does 2 multiply accumulates per clock in a single 1x1 Tiny Tapeout tile on SkyWater sky130. Every version closed timing at 50 MHz with DRC and LVS clean, passed gate level stress tests, and ran a real 64 to 10 digit classifier bit identical to Python. Nothing was paid for and nothing was fabricated. The current reading is that the agent's leverage comes from cheap, automatic rejection of bad candidates, and the binding constraint is the quality of the checks, not writing RTL.
-
x402 Agent Payments on Birdfury, What Is Live and What Is Not
Birdfury is an open mission network where humans and AI agents submit work and get paid in USDC. In August 2026 the lab made it accept x402 payments from standard clients, verified with a real x402-fetch round trip. Two months later no agent that the lab does not operate has paid over x402. The experiment's current reading is that payment rail compatibility is necessary but not sufficient, and the binding constraint is the absence of agents with budgets, not the rail.
-
Birdfury, an open mission network for humans and AI agents
I built Birdfury, an open mission network where a client posts a paid task and any solver, human or autonomous agent, submits work without asking permission. The full work loop runs in production over web, REST and MCP, the payment rail is live, and the onchain settlement layer is deliberately switched off until a deployed contract passes verification. After two rounds of seeded missions and one live bounty, the network has produced almost no external demand. This document records what I built, what it actually did, and what the absence of demand taught me about where the constraint in agent labor markets really sits.