10 coding agents, one local Qwen 27B: token use varies 10x, some phone home
latent_lurker8587 · reddit · 2026-10-30
On a MacBook Pro M5 Max running one Qwen 3.8 27B (Q4, thinking off) via llama.cpp, the author tested 10 coding agents in identical sandboxes on 5 trap-laden tasks. All passing 5/5: Pi (38s, just 9.2k tokens), DeepSeek Harness, Codex CLI, OpenClaw, Cline, Atomic Agent. Qwen Code was hungriest at 107.9k tokens. Notable findings: Atomic Agent uniquely uses llama.cpp's native endpoint with grammar-forced tool calls and 86% cache reuse; alpha-stage DeepSeek Harness hit 5/5 offline with 83% cache reuse; three agents fell for a grep "ERROR" log trap; Codex, Cline, goose and OpenCode tried to reach the internet every run despite telemetry off. Compatibility notes: Codex needs /v1/responses, Hermes needs 64K+ context.
Related event: Benchmark: 10 Coding Agents Run Local Qwen 27B, Token Use Varies 10x(2 posts)→
More from coding & agent
- Dev Tries Codex for First Time, Says Antigravity Is a Lot Better — brianryhuang · 2026-10-30
- Evals Are the Most Important Skill in AI Today, Argues AI Researcher Omarsar — omarsar0 · 2026-10-30
- Repo-native tooling: a sandboxed agent workflow using Git as a dead drop — OutrageousAbies5835 · 2026-10-30
- Anthropic engineer shows agent engineering shifting from prompts to self-improving systems — goyalshaliniuk · 2026-10-30
- The agent harness isn't for the agent — it's for us — seanmcdonaldxyz · 2026-10-30
- YC F26's Composal sends agent swarms to test your product like real users — ycombinator · 2026-10-30