Running coding agents on a 16GB MacBook for 3 months: the full setup
Puzzleheaded-Fox9821 · reddit · 2026-11-18
The author shares a 3-month setup for coding agents on a base M4 with 16GB RAM:
- Hard tasks go to Claude Code; small tasks use a local model in the pi coding agent, trimmed to 4 tools (read/bash/edit/write) — full tool sets and MCP schemas choke small models. Stripped down, it fixed 2 bugs and passed 3 failing tests in 72 seconds
- Magnitude as engine cut first-token latency from 6s to 2s, which matters since agents resend big prompts every turn
- 4B Gemma for reading/explaining code, 1.7B for summaries and commit messages (80 tok/s); thinking must be off on small models (73s reasoning with no answer vs 12s off)
- Dropped: local MCP servers, multi-agent, long sessions — new task, new session
More from coding & agent
- Proactive Agents That Schedule Their Own Work Hinge on Memory and UX, Says LangChain's Runkle — hwchase17 · 2026-11-18
- Cognition Launches Devin Bots: One Bot Leads a Team of Agents — aidangch · 2026-11-18
- Top token user vouches for Carbon: strong at xhigh for complex coding — olcan · 2026-11-18
- 2026's 10 Best AI Agent Frameworks: Agent Adoption Doubled, 90% Now Write Data — rseroter · 2026-11-18
- Replit Livestream With BabyAGI Creator Yohei Nakajima on Building Frontier Agents — ColleenMBrady · 2026-11-18
- fx: a tiny open-source Zig coding agent with a 6.33 MiB binary that runs in the browser — evilrabbit_ · 2026-11-18