Build AI agents that survive production.
Operator-grade teardowns and runnable systems for builders moving past demos. Read the work, or run it. Nothing here is theory - everything ships.
Notes
Teardowns, working notes, and field signals from building production agents. The public canon, updated when there is something to say.
Vault
Productised setups, templates, and operator kits. Built for Starkslab work first, packaged so you can run them today.
Currently featured
Rotates as new work shipsOpenClaw Source Code Structure: Gateway, Heartbeat, Skills, and Runtime Architecture
OpenClaw source code structure teardown: gateway architecture, heartbeat wakes, skills, writable memory, session routing, channel plugins, and runtime boundaries.
// excerpt from teardown
# operator router - entry
def route(intent, ctx):
lane = classify(intent)
if ctx.risk > "med":
return escalate(lane)
return dispatch(lane, ctx)01 / Recent notes
Browse all 57 notes02 / In the Vault
Browse all dropsClaude Agent Workspace (Anthropic Agent SDK)
You want a real agent workspace — not a chat tab. Something multi-workspace, tool-enabled, with files, repeatable runs, and BYOK keys per workspace — so you can build and ship agent workflows without duct-taping scripts together.
trustmrr-cli
You need verified startup revenue data — MRR, growth, churn, customer counts — but TrustMRR only has a web UI. No way to query it from your terminal or pipe it into agent workflows.
datafast-cli
DataFast has a clean analytics API, but there's no CLI. You can't check your site stats from the terminal, pipe them to scripts, or hand them to an AI agent as a tool. You're stuck in a browser dashboard.
Minimal Agent Framework (MAF)
Every AI agent framework is a maze of abstractions. You can't trace what happened, you can't replay a failed run, and when something breaks you're debugging the framework instead of your agent. You need something you can actually read.
From the doctrine
Agents do not need more theater. They need control loops that keep working after the demo.
Five principles for building agents that keep working after the first successful run.
Read the full doctrine