7 hard-won truths about working with coding agents
PawelHuryn · x · 2026-09-16
The author shares 7 lessons learned the hard way from working with coding agents:
- A goal alone isn't enough: explain the why and strategic context (market segments, value proposition, revenue model) with progressive disclosure so agents can make better autonomous decisions.
- Agents need something to review code against: documenting intent (e.g., roles and permissions) is critical — otherwise they can't tell a bug from a feature, and it's also a prerequisite for reliable tests.
- UX review is still weak: even with browser automation, dynamic elements, micro-interactions and resizing get missed. Asking agents to screenshot and inspect (exploratory testing) helps, but a human eye is still needed.
- Human review is only necessary when strategic context and documentation are missing; if you must review, review architecture and procedures — make releases more deterministic — not syntax.
- Letting Astra and Opus/Fable work together yields far better results — something neither OpenAI nor Anthropic will tell you.
More from coding & agent
- Mac beta gives Claude Code, Codex, Grok agents one shared home with unified memory — RileyRalmuto · 2026-09-16
- Burkov uses Codex, Claude Code and Grok interchangeably — and says that's bad news for all three — burkov · 2026-09-16
- Four LLM mayors, one deterministic city: open-source loop-design experiment wins CoreWeave Hacks top prize — hackgoofer · 2026-09-16
- ComfyUI Music Toolkit 3.0 adds YuE2-powered audio covers via SheetSage2 and ABC scores — Vivid_Promise1700 · 2026-09-16
- Steve Yegge: only Claude Fable is worth a damn, and it's too expensive — orchestrators won't work for a year — Steve_Yegge · 2026-09-16
- Running Grok Bot and Hermes as Project Managers, Then Stacking One Agent on Top of the Other — mazzaTalk · 2026-09-16