Open source · AI agent reliability · MMXXVI
Make the context answerable.
I study the moment an AI agent turns context into action: what evidence survived, what it actually used, and whether that evidence is still current before a tool touches the world.
Position
Long context is a capacity. Reliability is a chain of evidence.
ContextAdapt is an open-source project for identifying smaller context candidates without silently discarding action-critical evidence—and for checking goal, permission, and provenance again before execution.
Practical guide · English / 中文
Choosing context length for AI coding agents
Start with the evidence your task needs. Compare 200K and 1M, measure full cost, and verify the result. Eight practical questions, with examples and the limits of our published evidence.
Context troubleshooting · English / 中文
Why does Codex show 258K? Getting the 500K default to take effect
Separate the 258K display, the 500K default and the compaction threshold; check overrides and configure the window with documented settings.
New notes · English / 中文
Codex at 500K: more room, but check the state before writing
Start with the default, restore missing evidence, and check the current target, file state, and permission before writing.
计划没有错，世界已经变了：为什么长时程 Agent 的上下文问题最终会变成执行问题
一篇基于近期研究、受控实验与真实执行器集成的迷你综述：长上下文失败不只是遗忘，更可能是 Agent 在执行时仍依赖已经过期的事实与权限。
Latest field note
+22.4
percentage points over head-and-tail truncation in strict code-evidence retention.
95% CI · +2.4 to +40.6
How Much Code Context Does an AI Agent Need?
Can a coding agent use less context without losing the evidence needed to act? This note reports a controlled retention result, the baselines that remained competitive, and the eight tasks that still required full source.
Read the evidence →