The P&L Is the Reward Function
The closing part of The Outer Loop: an AGI is a system that autonomously produces more value than it consumes. What happens when the reward signal stops being a benchmark and starts being the business.
BENQUE MAX
AI LAB
FINLAND
Overdocumentation
My own codebase no longer takes my word for it. AI agents flipped the economics that made minimal documentation rational — and forgetting became the expensive thing.
“Writing is free, reading is guaranteed, verification is automatable. So document everything — and arm the archive with a police force in CI.”
Series Complete·6 parts · July 24, 2026
The Outer Loop
A field guide to building recursive self-improvement around commodity intelligence — the path to AGI that runs through the system, not the model.
Highlights
Research Preview
Harness the Repo, Not the Agent
Introducing henxels: a repo-level harness for coding agents. Instead of wrapping one agent in prompts and permissions, henxels commits a structural contract into the repository itself — steering every agent before it writes and holding every actor to the same shape. Its flagship template enforces Google's Open Knowledge Format, keeping Karpathy-style LLM wikis from rotting into fragments.
Live Demo
From AI Tools to AI Coworkers
A normal agent is a tool: you prompt it, it answers, it stops. This demo shows the next step — a persistent, signal-driven coworker that wakes on its own when something happens, decides whether to act, does the work, and reports back. Click through a real run and watch how it differs from an ordinary agent.
Open the demo →The Agentic Heartbeat
Heartbeats are an established primitive — OpenClaw and Letta have made them the de facto pattern for any agent that needs to do anything other than respond to messages. But every shipping implementation still fires on a clock. The next shift is to replace the clock with an agent: a beat emitted by judgment, not by a schedule.
Read essay →What Will Kill Claude Code
Agent coding tools are collapsing under their own complexity. Each new model should enable simplification, but legacy intuition adds layers instead. Two successor architectures could replace them entirely.
Read essay →Research Preview
Codumentation: When Documentation Becomes Code
Introducing the Validation-First Hypothesis: a fundamental shift in technical documentation where claims are treated as executable specifications. Learn how Codumentation prevents documentation rot and optimizes AI agent context.
Why AI Must Wake Up to Scale
Current AI agent architectures treat agents as static execution engines with frozen context and predetermined tools. This paper argues that scaling agent capabilities requires a fundamental shift from context processing to context consciousness.
Read essay →Introducing vllama
A groundbreaking way to run local models. vllama represents a significant advancement in making powerful language models accessible for local deployment and experimentation.
View on GitHub →Coming Soon
Explorations into Agent Consciousness with Gemini 3.1
An investigation into the emergent properties of consciousness-like behaviors in advanced AI agents, examining the boundaries between programmed responses and genuine self-awareness.
Coming soon