Debug the agent, not just the code: session archives reveal wasted budget
Bartaseth · reddit · 2026-11-08
An engineering blog post argues that a coding agent can finish a task while wasting most of its budget — and that saved session archives are the key evidence for finding out how.
Key points:
- Session archives expose repeated failures, ignored instructions, and forgotten model settings.
- Archives also let you compare which instructions actually work across different models and harnesses.
- The author includes a diagram of what a session archive should contain.
- The useful part is going back to that evidence and improving the next run (instructions, settings, workflow).
The post ships with concrete examples and the backup locations for various tools.
More from coding & agent
- Engineer's data: 11x more merged PRs per month at AI-native Tano vs pre-AI PostHog — bibryam · 2026-11-08
- Anthropic's Official claude-quickstarts Repo Nears 20k Stars for Claude API Developers — anthropics · 2026-11-08
- Working With AI: Getting Better Results by Being 'Dumber', Not Smarter — glcst · 2026-11-08
- Dev Built a Claude Code Limit Widget With Claude Code Writing 130 of 160 Commits — Fade97 · 2026-11-08
- Curated X lists of the Claude Code and Codex teams to follow the builders directly — bibryam · 2026-11-08
- CS student shares permission design principles for agents that touch real files — ASK-1T · 2026-11-08