Flue vs Sandcastle vs a custom coding agent harness

In this May 2026 walkthrough, Elie Steinbock compares Flue and Sandcastle with a custom TypeScript development harness he built for his company, Inbox Zero. He says the libraries did not support the Claude Max subscription route he wanted at the time, so he wrote his own orchestration to switch between Claude and Codex.

Flue examples cover model calls, JSON results and workflows defined through skills. He explains a Cloudflare R2-backed virtual filesystem, persisted sessions, restricted CLI access and remote sandboxes. A short just-bash example shows shell commands running against an in-memory filesystem.

Sandcastle receives a separate walkthrough of its Docker-based run command, repeated task execution, iteration counts and configuration options. The speaker explains Git worktree handling, copying an environment file and templates that combine implementation with review. These examples concern coding-agent orchestration, not offline model inference.

His custom harness retrieves a Linear issue, downloads its images, creates a worktree and installs dependencies with PNPM. It classifies the task, implements it with TDD, simplifies and reviews the code, creates a pull request, addresses review comments, sends a Slack update and cleans up. This is his company's development workflow; it is not a demonstration of Inbox Zero's email features. He values control over the steps and notes that some prompts still need refinement.