Shipping at Inference-Speed

Takeaways

• Peter Steinberger (Dec 2025): shipping speed now limited by inference time + hard thinking — most software is boring data-shuffling; start as CLI so agents can verify output.
• GPT-5 / Codex unlock: rarely reads code anymore — watches stream, knows architecture; important decisions are language/ecosystem and dependencies (TS web, Go CLI, Swift macOS/iOS).
• Codex vs Opus: Codex reads extensively before writing (slow start, fewer bad fixes); Opus eager on small edits, weaker on large refactors. Plan mode is a hack for older models — conversation + “build” replaces it.
• Workflow: 3–8 parallel projects, queue prompts in Codex, commit to main, cross-reference ../other-project, docs in docs/ + global AGENTS.MD, short prompts with screenshots for UI.
• Oracle tool: GPT-5 Pro CLI for stuck agents — less needed after GPT-5.2; knowledge cutoff (Aug vs Mar) matters for latest tools.
• Engineer for agents: codebases structured so agents work efficiently; “write docs to docs/*.md” and let model pick filenames.

Notes

Follow-up to Oct 2025 “Just talk to it” post. VibeTunnel zig rewrite (5h, one shot) as concrete capability benchmark. Clawd/OpenClaw ecosystem context. Default model: gpt-5.2-codex high.

Open questions

• Commit-to-main solo workflow — team adaptation patterns?
• When does not reading code become a liability for security/architecture debt?