Skip to content

SASE Blog

The SASE Blog is the publishing surface for essays about Structured Agentic Software Engineering (SASE): durable work units, orchestration, provider-independent workflows, review state, and the coordination layer around coding agents.

Start Here

Start with SASE: Structured Agentic Software Engineering, the launch post: why SASE wraps agent CLIs, how XPrompts work, what the ACE Agents tab changes, and how to install the tool.

If you want the practical path first, use Getting Started: Your First 15 Minutes: install, provider readiness, a safe first run, and the first visible agent record.

More posts are in the repository as drafts and will be published over time. The generated archive below lists only the entries included in the public site.

From Reading To Practice

  • Start with ACE, the Agentic Change Explorer, to learn the terminal interface for daily agent work.
  • Read the SDD flow to see how Spec-Driven Development turns plans and epics into phase work, while standalone task beads capture independent follow-ups.
  • Open the repository for source, issues, and implementation details.
  • 12 min read

SASE: Structured Agentic Software Engineering

The status quo is useful enough to be dangerous. Open a handful of terminal or tmux windows, run one coding-agent CLI in each, hand each one a scoped task, and hop between them as they finish. I think of this as the Boris Cherny method because Boris Cherny's Claude Code setup thread described running multiple Claude sessions in parallel, and Anthropic now documents parallel Claude Code sessions with worktrees as a normal workflow.

I automated my version of it with tmux_ai_window, a small script in my public dotfiles. One tmux binding opens a display-menu with claude, codex, agy, qwen, and opencode. Each installed CLI gets an accent color and a one-key shortcut; missing CLIs are greyed out. Choosing one opens a new window named ai, ai2, ai3, and so on, in the current pane's directory, with the relevant "yes, go do the work" flags already wired. When the CLI exits, the window closes and the remaining ai* windows are renumbered.

That got me a long way. It also made the missing layer painfully obvious.