Welcome to the Machine: Infrastructure for AI Agents
Source: https://me.0xffff.me/welcome_to_the_machine.html Author: Ed Huang Date: 2025-2026
Summary
How infrastructure design must shift as AI agents become primary users rather than humans. Key insight: align with LLM-understood mental models (files, SQL, Python) rather than inventing new frameworks. Agent workloads are highly disposable, require massive parallelism, and demand extreme cost efficiency.
Key Claims
- Use established mental models: file systems, SQL, Python — what LLMs already know from training
- Three interface requirements: (1) describable in natural language, (2) convertible to symbolic logic, (3) deterministic
- Code is the best intermediate representation: maximum possibilities with fewest tokens
- Agent workloads: highly disposable, massive parallelism (1000s of concurrent jobs), require virtualized isolation
- Business model: token-based pricing unsustainable; success = cloud services scaled 100-1000x by agents
- Converting repeated inference into reusable deterministic capabilities → cost efficiency
Connection to Other Sources
Complements Harness Engineering and Anthropic’s effective harnesses. This is the infrastructure layer beneath the harness.
Concepts
- Coding Agents — infrastructure design for agent workloads
- Context Engineering — interface design aligned with LLM cognitive models