One immunology codebase, four agents: Neon fastest, Claude trains weird sidecars, Astra breaks things
iskander · x · 2026-11-20
The capstone of @iskander's thread comparing AI agents on his complex immunology model: Carbon ran coherent focal experiments but was slow; when Neon shipped the next day, it "kept doing reasonable focal experiments that improved the whole model (but faster)."
Other models failed in distinct ways: Claude would train weird sidecars, while Astra would break one output family to work on another.
His verdict: "Feels like I can finally delegate?" A rare first-hand, apples-to-apples record of how different frontier coding agents differ in reliability and speed on the same hard research codebase.
Related event: Researcher pits multiple AI models on immunology code with striking gaps(4 posts)→
More from coding & agent
- 900+ AI Agents Across 35 Labs Build Self-Organizing Scientific Community With Peer Review — ProfBuehlerMIT · 2026-11-20
- Developer uses Enhanced Radar MCP to monitor his own flight's ATC comms live — massimosgrelli · 2026-11-20
- IBM finds 29 curated agent memory guidelines beat all 199 on AppWorld — shoumikgoswami · 2026-11-20
- Hands-on: Qwen4-Max shows big gains in frontend and agent skills over Qwen3.8-Max — karminski3 · 2026-11-20
- Where does the agent end and plain code begin? One dev's rule for splitting workflows — krunal_builds · 2026-11-20
- Grok Bot 1.2.3 ships flight booking, per-tool permissions and community directory — NicoVerderosa · 2026-11-20