Articles

Everyone says you don't write prompts anymore, you write loops — the practice now called loop engineering. It's true, and it's where the misunderstanding starts. The loop is a dozen lines; what it runs on is the harness — the goal it pursues, the checks it grades itself against, the documents that shape the tasks it writes for itself. Loop engineering is mostly harness engineering, and a loop is only as good as the harness around it.

2026-06-20 · agents, claude-code, harness, loops, autonomy, evals

A better model helps at the margin; the order-of-magnitude difference is the harness around it. But every harness is a bet against the model — half of it scaffolding you rent until the next release makes it dead weight, half of it judgment that is really just you, written down. A field guide to building both, and knowing which is which.

2026-06-02 · agents, claude-code, harness, context, skills, memory

x402's rails are built. Demand isn't. The gap tells you what kind of services x402 is actually for — and why most of them haven't been built.

2026-05-17 · agents, x402, micropayments, permissionless, zkstash

Notes from building realtime voice agents in 2026 — the shape of a Next.js + LangChain + OpenAI Realtime + Twilio stack, why most of the engineering lives in the integration layer, and why NAT is still a thing two decades on.

2026-05-07 · voice-agents, sip, telephony, ai-agents, llms, nat

The unsolved problems in agent memory — portability, ownership, cross-platform identity — are infrastructure problems that look identical to the ones Web3 wallets already solved.

2026-05-01 · agents, memory, web3, identity, zkstash

Lessons from a year running a multi-chain AI agent terminal across Solana and Base — what works, what doesn't, and where the real engineering effort goes once you hand an agent a wallet, a toolbox, and an audience.

2025-01-31 · agents, hiero, langgraph, production, web3