Coding agents open PRs that pass review, then go blind once changes hit the pipeline
sonichigo-1219 · reddit · 2026-11-12
A developer describes a recurring pattern: agents read the repo, edit a dozen files, and open PRs that pass review — but once changes enter CI/CD, the agent sees nothing. Failed builds, vulnerable dependencies, latency regressions, and incidents a week later all live in the delivery system, which the agent has no visibility into.
He raises two open questions:
- Context vs memory: context engineering decides what the agent knows for a task; memory persists across sessions. Fixing the post-merge blind spot looks like a context problem, not a memory problem.
- Permissions: an agent recommending a deploy shouldn't count as permission to deploy. He leans toward risk tiers determining how much autonomy a change gets.
He asks practitioners what their agents actually see after merge, and whether autonomy tiers exist in practice or everything is still human approval gates.
More from coding & agent
- Manus drops coding interviews, halves engineer hiring as PMs write code — bestvicam · 2026-11-12
- Aether cloud coding agent flexes sandbox CPU: 25.7 runs/s vs 19.1 and 15.3 rivals — Med1_Ai · 2026-11-12
- Fable Ultracode burns a 5-hour Max limit in 17 minutes by spawning 30 parallel researchers — teortaxesTex · 2026-11-12
- vLLM team open-sources OpenDirector, an agent-driven AI video creation workflow — vllm_project · 2026-11-12
- Leak: xAI Working on a Selection Box for Grok Bot to Comment on Files — nima_owji · 2026-11-12
- Coding Agent Tooling Layers: One Agent vs Many, Personal vs Org-Wide — jiayuan_jy · 2026-11-12