Coding Agents

Summary

Coding agents range from maximal platforms (Claude Code with 14,000 token prompts) to minimal harnesses (Pi with 4 tools and <1000 token prompts). The field is converging on Skills as the standard for packaging procedural knowledge with progressive disclosure.

Related Concepts

Concept	Summary
Pi Coding Agent	Minimal agent by Block: 4 tools, TypeScript extensions
Minimal Agent Harness	Smallest tool set; trust model's RL training
Agent Slop	Compound errors from high-velocity agents without friction
Agent Skills	Organized folders of procedural knowledge; progressive disclosure

Key Entities

Entity	Role
Mario Zechner	Pi creator; "slow the fuck down" plea
Barry Zhang	Anthropic engineer, co-creator of Agent Skills
Mahesh Murag	Anthropic engineer, co-creator of Agent Skills
Zubair Trabzada	Built GEO audit system with 12 Skills
Owain Lewis	Pi course author, extension builder
Andrej Karpathy	Coined "vibe coding", token shift observation

Key Sources

Source	Type	Date
Building Pi World Of Slop Mario Zechner	video	2026-04-17
Dont Build Agents Build Skills Anthropic	video	2025-12-09
Stop Building Agents Zubair Trabzada	video	2026-03-05
Pi Coding Agent Free Course Owain Lewis	video	2026-04-18

What's Missing

• Agent evaluation benchmarks beyond Terminal Bench
• Security analysis of autonomous coding agents
• Multi-agent collaboration patterns
• Agent-to-agent communication protocols
• Cost analysis: token consumption across agent architectures