The Hard Part of an Agent Loop Isn't the Loop

Yashas Gujjar — Wed, 08 Jul 2026 05:36:06 +0000

Originally posted on the Puzzle Quest Games blog.

In 2026, Boris Cherny, the creator of Claude Code at Anthropic, told an interviewer he'd stopped prompting the model by hand: "I don't prompt Claude anymore... My job is to write loops," (OfficeChai, 2026). For about a week my feeds treated that clip as proof loops are the next thing in AI coding. I'd been running one since March 2026, and the viral cut skipped the only part that's actually hard.

The Bootstrap Problem. I first ran into this pattern as a "Ralph Wiggum loop," before Cherny's clip made the rounds. Claude Code's /loop command and Copilot's "Automation" tab chase the same idea, but both are generic: wiring a real loop in Copilot CLI still means writing charters, skills, and routing from nothing, per project. That's boilerplate, the loop was supposed to remove it.

Squad is the bootstrap layer.

bradygaster / squad

Squad: AI agent teams for any project

Squad

English | 中文

Human-led AI agent teams for any project. One command. A team that helps you move faster with your code.

⚠️ Alpha Software — Squad is experimental. APIs and CLI commands may change between releases. We'll document breaking changes in CHANGELOG.md.

What is Squad?

Squad gives you a human-directed AI development team through GitHub Copilot. Describe what you're building. Get a team of specialists — frontend, backend, tester, lead — that live in your repo as files. They persist across sessions, learn your codebase, share decisions, and help you move faster without giving up oversight.

Squad is a productivity tool for humans, not a replacement for engineers, reviewers, or decision-makers. People stay accountable for priorities, approvals, and final changes; Squad helps with coordination, repetition, and parallel execution.

It's not a chatbot wearing hats. Each team member runs in its own context, reads only its own knowledge…

View on GitHub

A few commands scaffold an agent charter, then you hand each agent the skills it needs. A loop.md file is read at the start of every cycle, self-directed or off a GitHub issue queue, and agents file their own issues when they find work outside scope. Mine runs 30 minutes on, 10 off (the break matters: less than 10 and the next cycle collides with a subagent still finishing background work).

Where it falls short: Copilot-only (no good Claude Code/Codex port), and subsquads (parallel sub-teams) slowed my throughput instead of raising it.

Anyone else running agent loops against a GitHub issue queue instead of a fixed prompt? Curious what your routing rules look like.

Cleaner AI code isn't a smarter model

Yashas Gujjar — Wed, 01 Jul 2026 04:16:22 +0000

I built 5 iOS apps in 2 months with an agent loop. What keeps the code clean is deterministic tracks and hard gates, not the model. 10 years in AAA games taught me quality has to be mechanical.

Read the full stack →