Strip your coding agent to one bash tool and let it build its own memory system
jasonkneen · x · 2026-08-18
The author argues context loading (codebase, memory systems) matters far less than people think, and suggests a counterintuitive agent config:
- Give the coding agent a single tool (bash)
- Let it load skills on demand
- Ask it to design its own ideal memory system
- Allow it to spawn subagents with the same abilities
With a good model this yields fewer tokens, higher quality, and one-shot fixes—rarely exceeding 100k context. He claims most modern harnesses are 90% bloat, making agents more cautious, dumber, and wrong more often.
More from coding & agent
- How to integrate LLM into Flux Image-to-Image workflow — JmakMarshal · 2026-08-18
- Open Source End-to-End Stock Report System Using LSTM and Agentic AI — kmeanskaran · 2026-08-18
- poolside teases Model Factory updates: experiment API, trajectory store, new UI — dhruv2038 · 2026-08-18
- Claude writes a MacOS driver for an obscure Windows-only HP printer — fofrAI · 2026-08-18
- Building an AI with a brain you own using Markdown and 5 tools — dfinke · 2026-08-18
- Claude Code adds /design command for in-terminal UI mockups — The Decoder · 2026-08-18