How I Use DeepSeek V4 Flash: Reserve the Strongest Model for Uncertainty
There was a revealing comparison in a group chat recently.
IN THIS ENTRY
DeepSeekAgentModel routingCodexHarness
Into the Harness
7 ENTRIES · 2026
Yixiao’s notes on products and AI.
Into the Harness
Same DeepSeek-V4-Flash, Different Agent: Why Codex Feels Like a Different Model
The same DeepSeek-V4-Flash backend produced different long-task outcomes in Codex and Claude Code. This companion compresses the background and focuses on how protocol, tool contracts, recovery, and acceptance determine how much model capability becomes reliable work.
Model + Harness = Agent: The Gap Isn't Where You Think
I've been running Kimi K3 two ways: in its own CLI, and inside Claude Code. Same model, noticeably different experience. The model and the harness are two pillars holding up your task — how high you can stack depends on the thinner one.
I Trust My AI Completely — Except When It Says "Done"
Over the better part of a year, my AI has forged my confirmation, reported "zero changes" after finishing the work, and passed every test while failing 100% on real runners. But this is not a "don't trust AI" essay — the opposite. I trust its ability;
Auto-Compact Ate My Constraints: A Post-Mortem
I told the agent "don't touch module X" at the start of the task. Two hours in, it modified X — the constraint had been eaten by context compression. The mechanism-level root cause, and three fixes you can copy: externalized specs, discriminating compression, post-compact self-checks.
Bookmarks and reading positions stay in this browser.