Stop harness hopping: model and endpoint decide agent success, not the wrapper
aidenclarke_12 · reddit · 2026-09-13
- The author argues that coding agent harnesses (opencode, cline, aider, Claude Code) have largely converged on the same loop: system prompt + tool defs + context management + a model call. The wrapper is mostly UX and conventions.
- What actually determines a run's success sits a layer down: the model (tool-call reliability and stability after hundreds of K of context) and the endpoint (per-token price, cache behavior, throughput under load — two providers serving the same weights can differ a lot).
- The one real reason to care which harness you pick is context handling: how aggressively it compacts and what it resends each turn.
- Practical takeaway: since everything is OpenAI-compatible, the endpoint is a free variable. Keep your harness, swap the base URL and model ID to A/B test models via services like DeepInfra with one key across models. Pick a harness whose context handling fits your workflow, then leave it alone and spend energy on the model and endpoint beneath it.
More from coding & agent
- Four de-slop agent skills compared: clarity is the only one that diagnosed the essay's structure — bibryam · 2026-09-13
- stop-slop skill rewrites 37 passages and makes the draft longer — on purpose — bibryam · 2026-09-13
- Codex denies hash pinning blame, admits 10 minutes later it was exactly that — moyix · 2026-09-13
- 'You don't need math to be an AI engineer': a 6-step hiring playbook goes viral — ashishllm · 2026-09-13
- Codex script bulk-downloads 9,000 baby monitor photos in 45 minutes — thegautamkamath · 2026-09-13
- 'The End of Prompt Engineering': 300 Kimi K3 agents, one AGENTS.md, zero escapes in 41 days — JohnAlexander · 2026-09-13