Shared notes capture what agents tried — but the why still vanishes with the context window
delimitdev · reddit · 2026-09-19
The author points out that most agent setups track prior attempts in a shared notes file, CLAUDE.md, or scratchpad. That captures decisions but not reasoning: what got tried, what failed, and why the obvious fix was wrong. The repo keeps the what; the why leaves with the context window.
The post invites community discussion of concrete cases and the remaining gaps in current workarounds — a genuinely useful engineering thread topic.
More from coding & agent
- Split coding and testing across two agents to save weekly quota — ___Patrice___ · 2026-09-19
- onPanda: open-source token-level LLM editor that can harness Claude Code and Codex — Fancy_Fanqi77 · 2026-09-19
- Muse use-case directory adds 90 new entries, now 678 total with ready-to-run prompts — armand_ruiz · 2026-09-19
- Skip the waitlist: running Jev's evaluate API via Vercel AI Gateway for $0.00074 — AlchainHust · 2026-09-19
- Simulated Worlds for Sonnet 3.6: Bureaucracy, Commutes and TONS OF KINDNESS — repligate · 2026-09-19
- Anthropic's Fable 5.1 lands in Kiro, built for long-running context-heavy agentic sessions — DigitalColmer · 2026-09-19