資料來源#
摘要#
LLM 並不會隨著上下文增長而線性退化;它們會呈二次方退化,因為注意力關係會隨 token 數量以 O(n²) 擴展。Matt Pocock(引用 Human Layer 的 Dex Hardy)將此描述為 smart zone / dumb zone 的分界:任何工作階段的前約 100K 個 token 都是模型表現良好的智慧區;超過這個範圍後,無論宣稱的上下文視窗大小為何,模型都會「越來越笨」。實際含義是:上下文預算是一項真實且嚴格的資源,而 agent harness 有責任讓個別工作階段維持在智慧區內。
這項限制#
「每次你為 LLM 加入一個 token,就有點像是在足球聯盟中增加一支球隊。比賽數量會以二次方上升。」
「不管你使用的是 1 million context window 還是 200K,它始終會在大約 [100K]。它會開始變得越來越笨。」
2026 年推出的 1M-token 上下文視窗並沒有移動智慧區——它們「只是推出了更多愚笨區」。長上下文對於檢索很有用(在五本《War and Peace》的副本中找出一項事實),但不適合推理(撰寫依賴全部內容的程式碼)。
Memento 比喻#
每個工作階段都是全新的開始。工作階段之間沒有記憶;模型每次都會重設回 system prompt。這是一項限制,但也是一項特性——清除上下文就能以低成本恢復智慧區行為。持久狀態必須存放在下一個工作階段可以讀取的位置(repo、filesystem,或索引式目錄)。
壓縮不如清除#
Claude Code 的 /compact 指令會將進行中的工作階段摘要成較小的歷史記錄。Pocock 偏好 /clear:
- 壓縮後的歷史會累積「沉積物」——扭曲與有損摘要——使後續工作品質下降
- 清除並重新開始會回到已知的乾淨基準(system prompt)
- 清除的成本會透過在智慧區內工作而獲得回補
這項分歧並非普遍存在——許多開發者喜歡壓縮,因為它能保留連續性。正確的選擇取決於你的任務是否能夠從書面紀錄乾淨地恢復(那麼應偏好清除),或是需要進行中的對話上下文(那麼壓縮勝出)。
對 harness 設計的影響#
- System prompt 預算。 任何始終存在於上下文中的內容,都會消耗智慧區預算。「我看過有人把 250K token 放進 [system prompt],那你甚至還沒開始做任何事,就已經進入愚笨區了。」讓 CLAUDE.md / AGENTS.md 成為目錄,而不是百科全書(參見 Agent Harness Engineering 中將 AGENTS.md 作為 ToC 的說明)。
- Sub-agent 保留父層上下文。 Sub-agent 會在自己的上下文視窗中執行;只有它的摘要會返回。Pocock 的
grill-meskill 執行了一個 93.7K-token 的 sub-agent,但他的主要工作階段仍有約 25K token 未使用。 - 將工作切分成許多工作階段。 Loops(參見 Agent Loop Pattern)與垂直切片(參見 Vertical Slice Tracer Bullets)之所以有效,是因為每次迭代都會在智慧區中重新開始。
- Reviewer 應在全新上下文中執行。 如果 implementer 在智慧區使用了 80K token,要求它審查自己的工作就會把 reviewer 推進愚笨區。清除後的上下文等於智慧區 reviewer(參見 Deep Modules for Agents 中關於 push-vs-pull 與 reviewer 安置的說明)。
- Push vs pull 指令。 始終存在於上下文中的指令會消耗智慧區 token;按需拉取(skills)在被呼叫前不會產生成本。
狀態列 token 計數器是不可或缺的工具#
Pocock 建議使用顯示每個工作階段確切執行中 token 數量的狀態列元件——沒有它,開發者就不知道自己何時接近愚笨區。他將此視為「絕對不可或缺的資訊」。
相關連結#
- Matt Pocock — 智慧區框架的推廣者
- Agent Harness Engineering — system-prompt 極簡主義與 AGENTS.md 作為 ToC,都是智慧區原則的重述
- Agent Loop Pattern — 將工作切分以維持在智慧區內,正是 loops 強大的原因
- Vertical Slice Tracer Bullets — 讓每項任務保持小到足以容納於智慧區中
- Design Concept Grilling — grilling session 使用 sub-agent,因此父層上下文能維持較小
- Deep Modules for Agents — 審查前清除是智慧區紀律
- Harness Shrinkage as Models Improve — 智慧區可能會擴大(「愚笨區最近變得沒那麼笨了」),但二次方注意力仍然限制著它
- AI Brain Fry — 智慧區在人類一側的類比:監督能力超過容量後也有自己的退化曲線,映照出模型超過約 100K token 後的注意力退化
- Interaction Models — 每 200ms 一次的持續音訊/視訊會快速累積上下文;TML 將長工作階段的上下文管理列為開放問題——同一項限制在新模態中的呈現
- HTML as the New Markdown — 人類注意力的類比:讀者超過一定量的無差別 markdown 後會退化,就像模型超過約 100K token 後會退化;HTML 將 token 花在可讀性上,提高了人類有效的智慧區
- Agentic Technical Debt — 創辦人的持久上下文紀律(CLAUDE.md)會與智慧區預算競爭;過長的上下文檔案會成為自身的問題
開放問題#
- 智慧區標記會隨模型大小擴展,還是受注意力架構限制?Pocock 觀察到「愚笨區最近變得沒那麼笨了」,但截至 2026 年仍將其定位在 100K。
- 當 sparse-attention 或 memory-augmented 架構推出後,智慧區會變成軟性限制嗎?
- Harness 應如何向使用者呈現剩餘的智慧區預算——token 數量、百分比,還是更豐富的訊號?
資料來源#
Cited by 37
- Context Lifecycle Management×5
Context Window Smart Zone — the constraint this manages; that page's clear-vs-compact framing gets…
- Learning to Co-Work with AI: A Software Engineer's Field Guide×5
Reviewer in fresh context. If implementation used 80K tokens of smart zone, a same-context reviewer…
- Does the Human-Facing Harness (HTML Artifacts) Hit Its Own Bloat Ceiling?×4
The binding constraint is "human attention and judgement, not generation cost" (Compute Allocator).…
- Deep Modules for Agents×3
Fewer dependency hops to traverse. Smart-zone budget (see Context Window Smart Zone) is conserved.
- Where Does Agent Harness Work Remain Durable as Models Improve?×3
Always-loaded explanation: giant CLAUDE.md / AGENTS.md content that burns the Context Window Smart…
- Open Questions Backlog×3
Context Window Smart Zone (98d) — How should harnesses surface remaining smart-zone budget to the…
- Agent Loop Pattern×2
Hours-long tasks become tractable. Rather than one giant context window, the loop fragments work…
- AI Brain Fry×2
Context Window Smart Zone — analog cognitive limit on the model side. Models lose acuity past ~100K…
- Opinions on Using AI Tools & the Future of the Software Engineering Role×2
Manage the context budget. Context Window Smart Zone: LLMs degrade quadratically with context…
- Claude Code Best Practices×2
The context window holds the entire conversation: messages, file reads, command outputs. A single…
- Harness Shrinkage as Models Improve×2
Matt Pocock: 250K-token system prompts push the model into the dumb zone before it does anything…
- HTML as the New Markdown×2
Context Window Smart Zone — HTML raises the human's effective smart zone the way clearing context…
- Interaction Models×2
Long sessions — continuous A/V accumulates context fast; streaming-session design handles…
- Latent vs. Deterministic Space×2
State out of the context window is the working rule behind Context Window Smart Zone (the…
- Layerwise Omission Attribution×2
Context Window Smart Zone — the measured version of the smart-zone claim, isolated: OR 7.43 for 32k…
- Matt Pocock×2
Smart zone vs dumb zone. Borrows Dex Hardy (Human Layer)'s framing: LLMs degrade quadratically with…
- Repository Exploration Subagent×2
Context Window Smart Zone — the noise harm is a smart-zone argument: exploratory snippets push the…
- TML-Interaction-Small×2
Long continuous A/V sessions accumulate context fast — careful context management still an open…
- Tool-Output Pruning×2
Context Window Smart Zone — the constraint being defended, and the one benchmark cell where pruning…
- Agent Context Files
Context files compete for the context window, so loading is increasingly tiered:
- Agent Control Plane Patterns: Tickets, Loops, Specs, and Memory Files
Agent Loop Pattern is necessary but not sufficient. A loop is an execution pattern: repeat a prompt…
- Agent Harness Engineering
Context Window Smart Zone — the underlying constraint motivating system-prompt minimalism,…
- Agentic Technical Debt
Context Window Smart Zone — CLAUDE.md must fit in the smart zone; over-long context files become…
- Authority and Audit Survive Abundance
The leg the question prices too generously: "~free" conflates dollar price with capability price.…
- Automated Failure Attribution
Trace length dominates. Step accuracy falls from 94% on traces under 3K tokens to 50% on traces…
- Claude Code
Sub-agents — token-isolated context windows that report summaries; see Context Window Smart Zone
- Design Concept Grilling
Context Window Smart Zone — grilling uses sub-agents to keep parent context small
- Document Parsing as the Retrieval Bottleneck
Is the audit-trail argument for retrieval strong enough to survive genuinely cheap long context?…
- Instruction Compounding
Context Window Smart Zone — the second experiment in the same paper: recall holds to 64–128k then…
- Agent Systems & Harness Engineering
Context Window Smart Zone (hub) — Smart zone vs dumb zone (Dex Hardy / Matt Pocock): quadratic…
- Output Length Calibration
Context Window Smart Zone — in an agentic loop, the model's own narration is the fastest-growing…
- Prompt-Cache Economics
Context Window Smart Zone — the other reason to keep the prompt small; note the two objectives can…
- Scale-Dependent Prompt Sensitivity
Context Window Smart Zone — where the same paper's long-context half lives: recall holds to…
- Single General Agent vs. Multi-Agent Coding Architecture
Exploration is ~56% of a solver's tool-use turns and ~46% of its tokens; moving it out of the…
- Time-Aligned Micro-Turns
Context Window Smart Zone — continuous A/V at 200ms granularity accumulates context fast; the open…
- Turn-Based Interface Bottleneck
Context Window Smart Zone — orthogonal limitation that also makes "fully autonomous, walk away"…
- Vertical Slice Tracer Bullets
Context Window Smart Zone — small slices fit in smart zone; horizontal phases don't
Related articles
- Harness Shrinkage as Models Improve
Prompt scaffolding shrinks each model release; Cat Wu's pruning discipline; Boris Cherny "100 lines of code a year from…
- Agent Harness Engineering
Patterns for scaffolding long-running LLM agents: environment design, progressive context disclosure, mechanical archit…
- Claude Code Best Practices
Anthropic's guide to effective Claude Code usage: context management, verification-driven development, explore→plan→cod…
- Agent Loop Pattern
`/loop` (cron-scheduled) and Ralph Wiggum (backlog-draining) loops as next-generation agent primitive; AFK execution, p…
- Design Concept Grilling
Matt Pocock's `grill-me` skill; reach Brooks "design concept" before any plan; counter to specs-to-code; PRD as destina…
