The hosts share their personal coding-agent setups. One moved to Codex (GPT 5.5) over cost and found it stronger for long autonomous runs. Both use the Superpowers plugin (spec → plan → implement), which keeps their workflow portable across Claude Code, Codex, and Cursor. Planning: verbose technical plans blur what the human asked for with the agent’s assumptions, so keep them high-level and visual (HTML), with an in-browser feedback extension. For UI work, design first in Pencil.dev and make visual parity the definition of done. Concrete Claude Code settings (no-flicker, no default worktrees, custom git instructions, larger skill-listing budget, keep sessions longer, remote control on startup), custom skills (“ask good questions”, “linear ops”), a “context recap” output style, and a lid-closed keep-awake tool. Autonomy comes from cloud-first stacks with per-PR ephemeral environments (Pulumi/Kubernetes, MirrorD), a dedicated Chrome profile for the agent (Agent Browser CLI), CLI over MCP, nightly routines, and a secrets vault.
Show: langtalks
Concepts: coding-agent-workflow · harness-engineering · ai-verification · skill-engineering · agent-security · context-engineering