Build agents that survive contact with reality.

Most agents are workflows in costume. Reliability lives in the harness, the loop, and the gate, not the model. This lab publishes the evidence.

Start here Read the book The evidence

An engraved tightrope walker heading toward a brick-red sun

Most agents are workflows in costume. Reliability lives in the harness, the loop, and the gate, not the model. Earn the complexity. Evidence over hype.

+3.4pp Workflow beat agent 2.4× Agent cost 2.2× Agent latency See the evidence →

LAB-002 · LAB

The durable agent, and whether ADK 2.0's crash-resilient runtime earns its overhead

FN-004 · FIELD NOTE

Loop engineering is a 30-year-old loop with a new hashtag

FN-003 · FIELD NOTE

Anthropic just told you the harness is the product

Part I

Build

Tools, loops, multi-agent without theater.

4 chapters →

Part II

Judge

Eval, oversight, when not to use agents.

3 chapters →

Part III

Operate

Deploy, govern, secure, remember.

6 chapters →

The Bench

Read the argument, then run it. Interactive experiments in the browser, no setup. Start with Loop Lab: toggle the fences and watch an agent either delete production or get stopped.

Run the Loop Lab →

"Most agents are workflows in costume. Earn the complexity."

Read the manifesto →