Ditching agent frameworks made my agents noticeably better: eval and side-effect checks are the real work
Rbovzz · reddit · 2026-11-19
A Reddit user shares why dropping LangGraph/CrewAI-style frameworks for a plain loop (prompt → tool call → run → feed back) made their agents noticeably more reliable.
Key lessons:
- Evaluation is the real work: one good run means almost nothing. Running the same task 10 times and logging every tool call surfaces failures a single demo never shows.
- The dangerous step is the side effect, not the thinking: anything that books, sends, charges, or deletes needs a check immediately after execution, not just at the end—catching a bad booking midway is cheap, finding it tomorrow is not.
The author asks whether others prototype framework-free or go straight to LangGraph/CrewAI with logging on top.
More from coding & agent
- Qwen3.8-27B coding-agent-adapted model trends on Hugging Face — harness-hub · 2026-11-19
- Agno's agent compaction: summary as index, originals stay searchable — pritisinghhhh · 2026-11-19
- "Copy your AGENTS.md if AI coding left you zero problems to solve" — burny_tech · 2026-11-19
- Dev open-sources Traffic Engine: DOM playbook cache makes repeat browser agent sessions cost $0 tokens — ystudio3d · 2026-11-19
- Hugging Face team preps docs to build, package and distribute Helion kernels on the Hub — ariG23498 · 2026-11-19
- Developer Adds a Dynamic Island to His App in 30 Minutes With Just 3 Prompts, No Swift Needed — PrajwalTomar_ · 2026-11-19