{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-09-15T00:02:40.869Z","headline":"Agent 长任务上下文工程解析：用预算控制、压缩、todo-state 和记忆对抗上下文溢出与目标丢失","description":"文章解析 Agent harness 层应对长任务中上下文溢出与目标丢失的四类机制：上下文预算与卸载、压缩、todo-state 复述和跨会话记忆。","url":"https://www.aioga.com/news/cmtzezuk0077wrowyi7a4qy7v/","mainEntityOfPage":"https://www.aioga.com/news/cmtzezuk0077wrowyi7a4qy7v/","datePublished":"2026-09-13T05:56:03.000Z","dateModified":"2026-09-13T05:56:03.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://www.marktechpost.com/2026/09/12/context-engineering-inside-the-harness-4-mechanisms-that-beat-context-overflow-and-goal-loss-on-long-horizon-tasks","https://aihot.news/items/cmtzezuk0077wrowyi7a4qy7v"],"canonicalUrl":"https://www.aioga.com/news/cmtzezuk0077wrowyi7a4qy7v/","directAnswer":{"@type":"Answer","text":"MarkTechPost文章称，长任务中的Agent可能出现上下文溢出、目标丢失和状态维护不足。文章归纳了上下文预算与卸载、压缩、todo-state复述、跨会话记忆四类harness机制。","url":"https://www.aioga.com/news/cmtzezuk0077wrowyi7a4qy7v/","dateCreated":"2026-09-13T05:56:03.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"marktechpost.com source article","url":"https://www.marktechpost.com/2026/09/12/context-engineering-inside-the-harness-4-mechanisms-that-beat-context-overflow-and-goal-loss-on-long-horizon-tasks","datePublished":"2026-09-13T05:56:03.000Z","provider":{"@type":"Organization","name":"marktechpost.com","url":"https://www.marktechpost.com/2026/09/12/context-engineering-inside-the-harness-4-mechanisms-that-beat-context-overflow-and-goal-loss-on-long-horizon-tasks"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.news/items/cmtzezuk0077wrowyi7a4qy7v","datePublished":"2026-09-13T05:56:03.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.news/items/cmtzezuk0077wrowyi7a4qy7v"}}],"aggregationSource":"MarkTechPost（RSS）","originalPublisher":{"name":"marktechpost.com","url":"https://www.marktechpost.com/2026/09/12/context-engineering-inside-the-harness-4-mechanisms-that-beat-context-overflow-and-goal-loss-on-long-horizon-tasks"},"geoDeepAnswer":null,"article":{"id":"cmtzezuk0077wrowyi7a4qy7v","slug":"cmtzezuk0077wrowyi7a4qy7v","url":"https://www.aioga.com/news/cmtzezuk0077wrowyi7a4qy7v/","title":"Agent 长任务上下文工程解析：用预算控制、压缩、todo-state 和记忆对抗上下文溢出与目标丢失","title_en":"","summary":"文章解析 Agent harness 层应对长任务中上下文溢出与目标丢失的四类机制：上下文预算与卸载、压缩、todo-state 复述和跨会话记忆。","source":"MarkTechPost（RSS）","sourceUrl":"https://www.marktechpost.com/2026/09/12/context-engineering-inside-the-harness-4-mechanisms-that-beat-context-overflow-and-goal-loss-on-long-horizon-tasks","aiHotUrl":"https://aihot.news/items/cmtzezuk0077wrowyi7a4qy7v","publishedAt":"2026-09-13T05:56:03.000Z","category":"行业动态","score":72,"selected":true,"articleBody":["An agent, in its simplest form, is an LLM calling tools in a loop. That loop works for short jobs. Give it a task that runs for an hour and 200 tool calls, and it breaks in 2 predictable ways. The AWS Samples design guide for autonomous cloud coding agents：https://aws-samples.github.io/sample-autonomous-cloud-coding-agents/design/agent-harness names them directly: shallow agents suffer from context overflow, get distracted (goal loss), and do not maintain state over long periods. The layer that fixes this is not the model. It is the harness, which AWS describes as managing everything but the model.","This article opens up that layer. Compaction, memory strategy, context budgeting, and todo-state are the machinery that turns a shallow loop into a deep agent. We look at how LangChain Deep Agents：https://www.langchain.com/blog/context-management-for-deepagents, Claude Code：https://code.claude.com/docs/en/context-window, Manus：https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus, OpenAI Codex：https://developers.openai.com/api/docs/guides/compaction, and Amazon Bedrock AgentCore：https://aws.amazon.com/blogs/machine-learning/build-context-rich-research-agents-with-deep-agents-and-bedrock-agentcore/ implement each one, with the actual thresholds they ship.","The obvious fix is a larger context window. The evidence says it helps less than expected. Chroma’s Context Rot report：https://www.trychroma.com/research/context-rot evaluated 18 LLMs, including GPT-4.1, Claude 4, Gemini 2.5, and Qwen3, and found that performance grows increasingly unreliable as input length grows, even on simple retrieval tasks. Anthropic’s context engineering guide：https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents explains the mechanism: attention creates n² pairwise relationships for n tokens, so every added token depletes a finite “attention budget.” Context is a resource with diminishing returns, not a bucket.","For an agent loop, this is worse than it sounds. Manus reports that a typical task needs around 50 tool calls, and that the input-to-output token ratio runs near 100:1. Each observation lands in context and stays there. The original instruction drifts toward the middle of the window, which is exactly where recall degrades. Goal loss is not only a model bug. It is the expected outcome of an unmanaged context on a long enough task.","The first job of a harness is deciding what never enters the window at all. Deep Agents ships 2 offloading rules with hard numbers. When a tool response exceeds 20,000 tokens, it is written to the filesystem and replaced with a file path plus a preview of the first 10 lines. When session context crosses 85% of the model’s window, older write and edit tool calls, whose full file contents already live on disk, are truncated to a pointer. Only after offloading runs out of room does the harness fall back to summarization.","Claude Code applies the same budgeting to what loads before the first prompt. Auto memory is capped at the first 200 lines or 25KB. MCP tool schemas stay deferred by default, with only tool names listed, and full schemas load on demand via tool search. After compaction, any re-read file over 5,000 tokens comes back as a path reference rather than content. The context window simulation in the Claude Code docs：https://code.claude.com/docs/en/context-window makes the payoff concrete: a research subagent reads 6,100 tokens of files and returns a 420-token result to the parent.","That subagent pattern is budgeting at the architecture level. Anthropic’s guide notes that each subagent may burn tens of thousands of tokens exploring, but returns a distilled summary, often 1,000 to 2,000 tokens. The AWS AgentCore walkthrough builds exactly this: a coordinator spawns 3 browser subagents in parallel, each in its own MicroVM, and an analyst subagent receives only their structured findings. AWS reports a 4 to 6 minute expected runtime, and notes that sequential processing would take up to 3x longer.","When offloading is not enough, the harness summarizes. Compaction is the practice of taking a conversation nearing the window limit, summarizing it, and reinitiating a new context with the summary. It is also where goal loss most often happens, because a lossy summary can drop the one constraint that mattered.","The implementations differ in what they promise to keep. Claude Code’s compaction prompt preserves architectural decisions, unresolved bugs, and implementation details while discarding redundant tool outputs. Right after compaction it re-reads up to 5 of the files modified most recently, reloads the rules matching those files, and re-injects invoked skill bodies, capped at 5,000 tokens per skill and 25,000 total. The docs are explicit that detailed instructions from early in the conversation may be lost, which is why persistent rules belong in the project-root CLAUDE.md, which is re-injected from disk. Users can steer the pass with /compact focus on the auth bug fix or move the trigger point with /autocompact .","Deep Agents made goal preservation a structural feature. Its summary is a structured document with dedicated fields for session intent, artifacts created, and next steps. The LangChain team added those fields after forced-summarization experiments showed the change improved performance. The full original transcript is also written to the filesystem, so a fact that was summarized away can be recovered by read_file later.","Compaction has moved into the API layer too. OpenAI’s Responses API offers server-side compaction via context_management with a compact_threshold , plus a standalone /responses/compact endpoint that returns a compacted context window containing an opaque encrypted compaction item; OpenAI instructs developers to pass that returned window unchanged into the next call. OpenAI says Codex relies on this mechanism：https://openai.com/index/equip-responses-api-computer-environment/ to sustain long-running coding tasks. The Claude Developer Platform：https://platform.claude.com/docs/en/build-with-claude/compaction exposes a compact_20260112 context-management edit with custom instructions and a pause_after_compaction option for inserting content before the model continues. When you write custom instructions there, they replace the default prompt entirely, so a compaction prompt is a real engineering artifact, not a setting.","Compaction protects the goal at the moment of summarization. Todo-state protects it on every turn in between. Manus described the trick plainly: its agent creates a todo.md and rewrites it step by step, checking items off. Rewriting the list recites the objectives into the end of the context, pushing the global plan into the model’s recent attention span and reducing “lost in the middle” drift. No architecture change is required. It is natural language used to bias the model’s own attention.","The evidence on todo-state is not one-sided. Deep Agents shipped a write_todos tool by default until v0.7 in July 2026：https://www.langchain.com/blog/deep-agents-v0-7, when LangChain made TodoListMiddleware opt-in after its evals across 3 task categories showed slightly better reward and lower cost with todos disabled. LangChain still recommends turning it back on for long multi-step tasks, less capable models, and UIs that show progress. Claude Code keeps a todo list and re-injects the plan written in plan mode from disk after compaction. Anthropic’s guide calls the general pattern structured note-taking: the agent writes a NOTES.md or TODO file outside the window and reloads it. Its Claude Plays Pokémon example maintained tallies across thousands of game steps, then read its own notes after each context reset and resumed multi-hour sequences.","The pattern behind all of these is that the goal exists as a mutable artifact, not only as a message in history. Messages age and get summarized. A file that is rewritten every few turns is always recent, always short, and survives any reset. Whether that is worth its per-turn token cost depends on the model and the task length, which is exactly what the Deep Agents evals measured.","The last piece is what persists after the task ends. Claude Code re-injects the project-root CLAUDE.md and auto memory from disk after every compaction. AgentCore Memory stores events and runs configured extraction strategies in the background, so a coordinator can call a recall tool on the next run instead of re-researching. AWS warns that without at least 1 extraction strategy configured, raw events are stored but nothing is extracted for retrieval. Anthropic’s file-based memory tool serves the same purpose on the Claude platform.","The limitation is that persistent context is not free. The ETH Zurich study we covered in February：https://www.marktechpost.com/2026/02/25/new-eth-zurich-study-proves-your-ai-coding-agents-are-failing-because-your-agents-md-files-are-too-detailed/ found that repository context files like AGENTS.md do not generally improve task success while raising inference cost: LLM-generated files increased cost by 20% and 23% on the 2 benchmarks, and developer-committed files by up to 19%. Memory that reloads every session is a standing tax on the attention budget. The Claude Code docs give the matching advice: keep CLAUDE.md under 200 lines and move reference material into skills or path-scoped rules that load only when needed.","The simulator below runs a 60-step migration task through a 200K token window. Toggle the 4 mechanisms, set the compaction trigger, and press Run. With everything off, the window overflows before the task is half done. With offloading, compaction, todo recitation, and subagent delegation on, the same task finishes with the goal still in recent attention. Token counts are illustrative; the thresholds match Deep Agents defaults.","Context management is only useful if the agent can still finish the task and recover details it no longer sees. LangChain maintains targeted evals for exactly this: tests that trigger summarization mid-task and check whether the agent continues toward its objective, and needle-in-a-haystack cases where a fact is summarized away and must be recovered through filesystem search. To generate enough events to compare prompt variants, the team triggers summarization at 10 to 20% of the window instead of the 85% default, and used a 25% trigger with Claude Sonnet 4.5 on terminal-bench-2 to study the effect.","The failure to watch for, in LangChain’s view, is goal drift: an agent that asks for clarification right after a summary, or wrongly declares the task complete. AgentCore Evaluations ships a goal success rate evaluator that can score the same traces. If you run a harness and have not forced a compaction in a test, you do not yet know what your summary prompt drops.","Practitioner-first AI/ML news and analysis, read by 1M+ developers and researchers every month."],"articleImages":[{"sourceUrl":"https://www.marktechpost.com/wp-content/uploads/2026/09/blog123-1-1-100x70.png","alt":"Implementation of Machine Learning Workflows with NVIDIA cuML, RAPIDS, GPU Benchmarking, Explainability, Clustering, and Model Inference","afterParagraph":18,"url":"/media/articles/cmtzezuk0077wrowyi7a4qy7v/f9a6bb8b9e03fb68.png"},{"sourceUrl":"https://www.marktechpost.com/wp-content/uploads/2026/09/blog111-5-100x70.png","alt":"Can LLMs Engineer Their Own Agent Harness? ByteDance Seed's HarnessDev Says Only 34 of 64 Changes Generalize","afterParagraph":18,"url":"/media/articles/cmtzezuk0077wrowyi7a4qy7v/60cd6c8dfd7a6b80.webp"},{"sourceUrl":"https://www.marktechpost.com/wp-content/uploads/2026/09/blog111-3-100x70.png","alt":"Cohere Releases North Small Translate","afterParagraph":18,"url":"/media/articles/cmtzezuk0077wrowyi7a4qy7v/dd3d62e2d8cf9fe8.webp"}],"mediaStatus":"ok","articleBodyZh":["一个代理，在最简单的形式下，就是一个循环调用工具的 LLM。这个循环适用于短任务。给它一个运行一小时并调用200次工具的任务，它会以两种可预测的方式崩溃。AWS Samples 自主云编码代理设计指南：https://aws-samples.github.io/sample-autonomous-cloud-coding-agents/design/agent-harness 直接提到了它们：浅层代理会遭遇上下文溢出、分心（目标丢失），并且无法在长时间内维持状态。解决这一问题的层不是模型，而是 AWS 所描述的管理除模型之外一切的外壳（harness）。","本文揭示了这一层。压缩、内存策略、上下文预算和待办状态是将浅层循环转变为深度代理的机制。我们研究 LangChain 深度代理：https://www.langchain.com/blog/context-management-for-deepagents、Claude Code：https://code.claude.com/docs/en/context-window、Manus：https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus、OpenAI Codex：https://developers.openai.com/api/docs/guides/compaction 和 Amazon Bedrock AgentCore：https://aws.amazon.com/blogs/machine-learning/build-context-rich-research-agents-with-deep-agents-and-bedrock-agentcore 如何实现每一个机制，以及它们实际使用的阈值。","显而易见的解决方法是更大的上下文窗口。证据显示其帮助没有预期的那么大。Chroma 的上下文衰减报告：https://www.trychroma.com/research/context-rot 对包括 GPT-4.1、Claude 4、Gemini 2.5 和 Qwen3 在内的18个 LLM 进行了评估，发现随着输入长度的增加，即使是简单的检索任务，性能也变得越来越不可靠。Anthropic 的上下文工程指南：https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents 解释了机制：注意力为 n 个标记创建 n² 的成对关系，因此每增加一个标记就会消耗有限的“注意力预算”。上下文是一种收益递减的资源，而不是一个桶。","对于一个代理循环，这比听起来更糟。Manus 报告说，一个典型任务大约需要 50 次工具调用，而且输入到输出的 token 比例接近 100:1。每个观察都会进入上下文并保持在那里。原始指令会漂移到窗口的中间位置，正是在这里回忆会下降。目标丢失不仅仅是模型的错误，这是在任务足够长时未管理上下文的预期结果。","一个挂载系统的首要任务是决定哪些内容永远不进入窗口。Deep Agents 配备了两个带有硬性数字的卸载规则。当工具响应超过 20,000 个 token 时，它会被写入文件系统，并替换为文件路径加前 10 行的预览。当会话上下文超过模型窗口的 85% 时，较早的写入和编辑工具调用（其完整文件内容已经存储在磁盘上）将被截断为指针。只有在卸载空间用尽后，挂载系统才会使用摘要作为回退方案。","Claude Code 对第一次提示前加载的内容也采取了同样的预算管理。自动内存限制在前 200 行或 25KB。MCP 工具模式默认保持延迟加载，仅列出工具名称，完整模式通过工具搜索按需加载。经过压缩后，任何超过 5,000 个 token 的重读文件会以路径引用返回，而不是内容。Claude Code 文档中的上下文窗口模拟：https://code.claude.com/docs/en/context-window 使收益具体化：一个研究子代理读取了 6,100 个 token 的文件，并向父代理返回一个 420 token 的结果。","这种子代理模式是在架构层面的预算管理。Anthropic 的指南指出，每个子代理可能会消耗数万个 token 进行探索，但返回的是精炼摘要，通常为 1,000 到 2,000 个 token。AWS AgentCore 演练构建了完全相同的模式：一个协调员并行生成 3 个浏览器子代理，每个子代理在自己的 MicroVM 中运行，而分析子代理只接收它们的结构化发现。AWS 报告预期运行时间为 4 到 6 分钟，并指出顺序处理可能需要长达 3 倍的时间。","当卸载不够时，系统会进行摘要。压缩是指将接近窗口限制的对话进行总结，并用该摘要重新启动一个新上下文的做法。这也是目标最容易丢失的地方，因为有损摘要可能会丢掉唯一重要的约束条件。","各实现方式在承诺保留的内容上有所不同。Claude Code 的压缩提示会保留架构决策、未解决的 bug 和实现细节，同时丢弃冗余的工具输出。压缩完成后，它会重新读取最近修改的最多 5 个文件，重新加载与这些文件匹配的规则，并重新注入已调用的技能内容，每项技能最多 5,000 个 token，总计 25,000 token。文档明确指出，对话早期的详细指令可能会丢失，这就是为什么持久性规则应放在项目根目录下的 CLAUDE.md 中，并从磁盘重新注入。用户可以通过 /compact 指令聚焦于授权 bug 修复，或通过 /autocompact 移动触发点。","Deep Agents 将目标保留作为结构特性。它的摘要是一个结构化文档，具有专门字段用于表示会话意图、创建的工件和下一步计划。LangChain 团队在强制摘要实验后增加了这些字段，结果显示这一更改提高了性能。完整的原始记录也会写入文件系统，因此被摘要掉的事实可以在稍后通过 read_file 恢复。","压缩（Compaction）也已进入 API 层。OpenAI 的 Responses API 通过 context_management 提供服务器端压缩，设置 compact_threshold，并且提供一个独立的 /responses/compact 端点，该端点返回包含不透明加密压缩项的压缩上下文窗口；OpenAI 指导开发者在下一次调用中不更改地传入返回的窗口。OpenAI 表示 Codex 依赖此机制：https://openai.com/index/equip-responses-api-computer-environment/ 以维持长时间运行的编码任务。Claude 开发者平台：https://platform.claude.com/docs/en/build-with-claude/compaction 提供了 compact_20260112 context-management 编辑，可包含自定义指令和 pause_after_compaction 选项，用于在模型继续之前插入内容。当你在此编写自定义指令时，它们会完全替代默认提示，因此压缩提示是真正的工程产物，而不是一个设置。","压缩在总结瞬间保护目标。Todo-state 在每个中间步骤保护它。Manus 简单地描述了这个技巧：其代理创建一个 todo.md 并逐步重写，逐项勾选。重写列表将目标复述到上下文末尾，将全局计划推进到模型的近期注意范围内，减少“中间丢失”的偏移。无需改变架构。这是使用自然语言来引导模型自身注意力的方式。","关于 todo-state 的证据并非单方面。Deep Agents 默认提供 write_todos 工具，直到 2026 年 7 月 v0.7 版本：https://www.langchain.com/blog/deep-agents-v0-7，当时 LangChain 在其对 3 类任务的评估中显示禁用 todos 略微提高奖励且降低成本后，将 TodoListMiddleware 设置为可选。LangChain 仍建议在长多步骤任务、低能力模型以及显示进度的 UI 上重新启用它。Claude Code 保持 todo 列表，并在压缩后从磁盘重新注入 plan 模式中编写的计划。Anthropic 的指南称这一通用模式为结构化笔记：代理在窗口外写 NOTES.md 或 TODO 文件并重新加载。在 Claude 玩《宝可梦》的示例中，它在数千步游戏中保持统计数据，然后在每次上下文重置后读取自己的笔记并继续多小时序列。","这些背后的模式是目标存在作为一个可变的工件，而不仅仅是历史中的一条信息。信息会随时间老化并被总结。每隔几轮重写一次的文件总是最新的，总是简短的，并且可以在任何重置后存活。它是否值得每轮的代币成本取决于模型和任务长度，这正是 Deep Agents 评估所测量的内容。","最后一部分是任务结束后仍然存在的内容。Claude Code 在每次压缩后会重新注入项目根目录的 CLAUDE.md 文件和磁盘上的自动记忆。AgentCore Memory 存储事件，并在后台运行配置的提取策略，因此协调器可以在下一次运行时调用回忆工具，而无需重新研究。AWS 警告称，如果至少没有配置一个提取策略，原始事件会被存储，但不会提取用于检索。Anthropic 的基于文件的记忆工具在 Claude 平台上也起到相同作用。","限制在于持久上下文并非免费的。我们在二月报道的苏黎世联邦理工学院的研究：https://www.marktechpost.com/2026/02/25/new-eth-zurich-study-proves-your-ai-coding-agents-are-failing-because-your-agents-md-files-are-too-detailed/ 发现，像 AGENTS.md 这样的仓库上下文文件通常不会提高任务成功率，同时会增加推理成本：LLM 生成的文件在两个基准测试中分别增加了 20% 和 23% 的成本，而开发者提交的文件增加了最多 19%。每次会话都重新加载的记忆是对注意力预算的长期负担。Claude Code 文档给出相应建议：将 CLAUDE.md 保持在 200 行以下，并将参考资料移动到技能或路径范围规则中，仅在需要时加载。","下面的模拟器通过 200K 代币窗口运行一个 60 步的迁移任务。切换 4 种机制，设置压缩触发，并按运行。关闭所有功能时，在任务完成一半之前窗口会溢出。打开卸载、压缩、待办事项背诵和子代理委派后，同一任务可以在目标仍处于近期注意中的情况下完成。代币计数仅供示例；阈值与 Deep Agents 默认值匹配。","上下文管理只有在代理仍能完成任务并恢复其不再看到的细节时才有用。LangChain 为此保持了针对性的评估：测试在任务中途触发总结，并检查代理是否继续朝着目标前进，以及在大海捞针的案例中，某个事实被总结掉后必须通过文件系统搜索恢复。为了生成足够的事件以比较提示变体，团队在窗口的 10% 到 20% 处触发总结，而非默认的 85%，并在 terminal-bench-2 上使用 Claude Sonnet 4.5 进行了 25% 的触发来研究效果。","在 LangChain 看来，失败的原因是未能关注目标漂移：代理在总结后立即请求澄清，或错误地宣布任务完成。AgentCore 评估提供了一个目标成功率评估器，可以对相同的跟踪进行评分。如果运行测试工具并且未在测试中强制压缩，你尚未知道总结提示会丢失哪些内容。","面向从业者的 AI/ML 新闻和分析，每月有超过 100 万开发者和研究人员阅读。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"MarkTechPost文章称，长任务中的Agent可能出现上下文溢出、目标丢失和状态维护不足。文章归纳了上下文预算与卸载、压缩、todo-state复述、跨会话记忆四类harness机制。","background":"材料将harness描述为模型之外的管理层，并介绍了Deep Agents、Claude Code、Manus、OpenAI Codex和Amazon Bedrock AgentCore对上下文管理的做法。文章还援引研究称，输入变长可能使模型表现更不稳定。","viewpoint":"Aioga判断：长任务能力不宜只按上下文窗口大小衡量。来源材料显示，工具结果如何进入上下文、旧信息如何保留，以及目标状态如何被重新呈现，都是Agent持续执行时值得关注的编辑观察点。","implications":"可能影响：Agent产品需要在信息完整性与上下文成本之间做管理。自动卸载和压缩可能减少窗口压力，但不代表关键信息一定被保留；产品评估也应关注长任务中的状态连续性，而不足以只看短任务表现。","nextStep":"后续观察：关注相关产品是否公开压缩触发条件、卸载规则、状态记录方式和跨会话记忆边界，并核对这些机制在不同任务长度与工具调用规模下的实际效果。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-09-13T06:47:30.308Z","sourceHash":"6cf564fb8aeb1003","review":{"approved":true,"groundedness":92,"clarity":84,"duplicationRisk":12,"blockingIssues":[],"notes":["“Aioga判断”属于观点标注，但名称与上下文衔接略显突兀，建议确认是否为预期署名或改为“编辑判断”。","“而不足以只看短任务表现”语法不完整，建议改为“而不应只看短任务表现”。","“自动卸载和压缩可能减少窗口压力，但不代表关键信息一定被保留”属于基于材料中压缩可能丢失关键信息的合理推论，可保留。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","editorial-labels","inference-boundary","low-source-overlap","no-html","independent-ai-review"]}},"tags":["行业动态","MarkTechPost（RSS）"],"translations":{"zh-CN":{"title":"Agent 长任务上下文工程解析：用预算控制、压缩、todo-state 和记忆对抗上下文溢出与目标丢失","summary":"文章解析 Agent harness 层应对长任务中上下文溢出与目标丢失的四类机制：上下文预算与卸载、压缩、todo-state 复述和跨会话记忆。","category":"行业动态","source":"marktechpost.com","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Agent 长任务上下文工程解析：用预算控制、压缩、todo-state 和记忆对抗上下文溢出与目标丢失 - Aioga AI资讯","description":"文章解析 Agent harness 层应对长任务中上下文溢出与目标丢失的四类机制：上下文预算与卸载、压缩、todo-state 复述和跨会话记忆。","url":"https://www.aioga.com/news/cmtzezuk0077wrowyi7a4qy7v/","articleBody":["一个代理，在最简单的形式下，就是一个循环调用工具的 LLM。这个循环适用于短任务。给它一个运行一小时并调用200次工具的任务，它会以两种可预测的方式崩溃。AWS Samples 自主云编码代理设计指南：https://aws-samples.github.io/sample-autonomous-cloud-coding-agents/design/agent-harness 直接提到了它们：浅层代理会遭遇上下文溢出、分心（目标丢失），并且无法在长时间内维持状态。解决这一问题的层不是模型，而是 AWS 所描述的管理除模型之外一切的外壳（harness）。","本文揭示了这一层。压缩、内存策略、上下文预算和待办状态是将浅层循环转变为深度代理的机制。我们研究 LangChain 深度代理：https://www.langchain.com/blog/context-management-for-deepagents、Claude Code：https://code.claude.com/docs/en/context-window、Manus：https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus、OpenAI Codex：https://developers.openai.com/api/docs/guides/compaction 和 Amazon Bedrock AgentCore：https://aws.amazon.com/blogs/machine-learning/build-context-rich-research-agents-with-deep-agents-and-bedrock-agentcore 如何实现每一个机制，以及它们实际使用的阈值。","显而易见的解决方法是更大的上下文窗口。证据显示其帮助没有预期的那么大。Chroma 的上下文衰减报告：https://www.trychroma.com/research/context-rot 对包括 GPT-4.1、Claude 4、Gemini 2.5 和 Qwen3 在内的18个 LLM 进行了评估，发现随着输入长度的增加，即使是简单的检索任务，性能也变得越来越不可靠。Anthropic 的上下文工程指南：https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents 解释了机制：注意力为 n 个标记创建 n² 的成对关系，因此每增加一个标记就会消耗有限的“注意力预算”。上下文是一种收益递减的资源，而不是一个桶。","对于一个代理循环，这比听起来更糟。Manus 报告说，一个典型任务大约需要 50 次工具调用，而且输入到输出的 token 比例接近 100:1。每个观察都会进入上下文并保持在那里。原始指令会漂移到窗口的中间位置，正是在这里回忆会下降。目标丢失不仅仅是模型的错误，这是在任务足够长时未管理上下文的预期结果。","一个挂载系统的首要任务是决定哪些内容永远不进入窗口。Deep Agents 配备了两个带有硬性数字的卸载规则。当工具响应超过 20,000 个 token 时，它会被写入文件系统，并替换为文件路径加前 10 行的预览。当会话上下文超过模型窗口的 85% 时，较早的写入和编辑工具调用（其完整文件内容已经存储在磁盘上）将被截断为指针。只有在卸载空间用尽后，挂载系统才会使用摘要作为回退方案。","Claude Code 对第一次提示前加载的内容也采取了同样的预算管理。自动内存限制在前 200 行或 25KB。MCP 工具模式默认保持延迟加载，仅列出工具名称，完整模式通过工具搜索按需加载。经过压缩后，任何超过 5,000 个 token 的重读文件会以路径引用返回，而不是内容。Claude Code 文档中的上下文窗口模拟：https://code.claude.com/docs/en/context-window 使收益具体化：一个研究子代理读取了 6,100 个 token 的文件，并向父代理返回一个 420 token 的结果。","这种子代理模式是在架构层面的预算管理。Anthropic 的指南指出，每个子代理可能会消耗数万个 token 进行探索，但返回的是精炼摘要，通常为 1,000 到 2,000 个 token。AWS AgentCore 演练构建了完全相同的模式：一个协调员并行生成 3 个浏览器子代理，每个子代理在自己的 MicroVM 中运行，而分析子代理只接收它们的结构化发现。AWS 报告预期运行时间为 4 到 6 分钟，并指出顺序处理可能需要长达 3 倍的时间。","当卸载不够时，系统会进行摘要。压缩是指将接近窗口限制的对话进行总结，并用该摘要重新启动一个新上下文的做法。这也是目标最容易丢失的地方，因为有损摘要可能会丢掉唯一重要的约束条件。","各实现方式在承诺保留的内容上有所不同。Claude Code 的压缩提示会保留架构决策、未解决的 bug 和实现细节，同时丢弃冗余的工具输出。压缩完成后，它会重新读取最近修改的最多 5 个文件，重新加载与这些文件匹配的规则，并重新注入已调用的技能内容，每项技能最多 5,000 个 token，总计 25,000 token。文档明确指出，对话早期的详细指令可能会丢失，这就是为什么持久性规则应放在项目根目录下的 CLAUDE.md 中，并从磁盘重新注入。用户可以通过 /compact 指令聚焦于授权 bug 修复，或通过 /autocompact 移动触发点。","Deep Agents 将目标保留作为结构特性。它的摘要是一个结构化文档，具有专门字段用于表示会话意图、创建的工件和下一步计划。LangChain 团队在强制摘要实验后增加了这些字段，结果显示这一更改提高了性能。完整的原始记录也会写入文件系统，因此被摘要掉的事实可以在稍后通过 read_file 恢复。","压缩（Compaction）也已进入 API 层。OpenAI 的 Responses API 通过 context_management 提供服务器端压缩，设置 compact_threshold，并且提供一个独立的 /responses/compact 端点，该端点返回包含不透明加密压缩项的压缩上下文窗口；OpenAI 指导开发者在下一次调用中不更改地传入返回的窗口。OpenAI 表示 Codex 依赖此机制：https://openai.com/index/equip-responses-api-computer-environment/ 以维持长时间运行的编码任务。Claude 开发者平台：https://platform.claude.com/docs/en/build-with-claude/compaction 提供了 compact_20260112 context-management 编辑，可包含自定义指令和 pause_after_compaction 选项，用于在模型继续之前插入内容。当你在此编写自定义指令时，它们会完全替代默认提示，因此压缩提示是真正的工程产物，而不是一个设置。","压缩在总结瞬间保护目标。Todo-state 在每个中间步骤保护它。Manus 简单地描述了这个技巧：其代理创建一个 todo.md 并逐步重写，逐项勾选。重写列表将目标复述到上下文末尾，将全局计划推进到模型的近期注意范围内，减少“中间丢失”的偏移。无需改变架构。这是使用自然语言来引导模型自身注意力的方式。","关于 todo-state 的证据并非单方面。Deep Agents 默认提供 write_todos 工具，直到 2026 年 7 月 v0.7 版本：https://www.langchain.com/blog/deep-agents-v0-7，当时 LangChain 在其对 3 类任务的评估中显示禁用 todos 略微提高奖励且降低成本后，将 TodoListMiddleware 设置为可选。LangChain 仍建议在长多步骤任务、低能力模型以及显示进度的 UI 上重新启用它。Claude Code 保持 todo 列表，并在压缩后从磁盘重新注入 plan 模式中编写的计划。Anthropic 的指南称这一通用模式为结构化笔记：代理在窗口外写 NOTES.md 或 TODO 文件并重新加载。在 Claude 玩《宝可梦》的示例中，它在数千步游戏中保持统计数据，然后在每次上下文重置后读取自己的笔记并继续多小时序列。","这些背后的模式是目标存在作为一个可变的工件，而不仅仅是历史中的一条信息。信息会随时间老化并被总结。每隔几轮重写一次的文件总是最新的，总是简短的，并且可以在任何重置后存活。它是否值得每轮的代币成本取决于模型和任务长度，这正是 Deep Agents 评估所测量的内容。","最后一部分是任务结束后仍然存在的内容。Claude Code 在每次压缩后会重新注入项目根目录的 CLAUDE.md 文件和磁盘上的自动记忆。AgentCore Memory 存储事件，并在后台运行配置的提取策略，因此协调器可以在下一次运行时调用回忆工具，而无需重新研究。AWS 警告称，如果至少没有配置一个提取策略，原始事件会被存储，但不会提取用于检索。Anthropic 的基于文件的记忆工具在 Claude 平台上也起到相同作用。","限制在于持久上下文并非免费的。我们在二月报道的苏黎世联邦理工学院的研究：https://www.marktechpost.com/2026/02/25/new-eth-zurich-study-proves-your-ai-coding-agents-are-failing-because-your-agents-md-files-are-too-detailed/ 发现，像 AGENTS.md 这样的仓库上下文文件通常不会提高任务成功率，同时会增加推理成本：LLM 生成的文件在两个基准测试中分别增加了 20% 和 23% 的成本，而开发者提交的文件增加了最多 19%。每次会话都重新加载的记忆是对注意力预算的长期负担。Claude Code 文档给出相应建议：将 CLAUDE.md 保持在 200 行以下，并将参考资料移动到技能或路径范围规则中，仅在需要时加载。","下面的模拟器通过 200K 代币窗口运行一个 60 步的迁移任务。切换 4 种机制，设置压缩触发，并按运行。关闭所有功能时，在任务完成一半之前窗口会溢出。打开卸载、压缩、待办事项背诵和子代理委派后，同一任务可以在目标仍处于近期注意中的情况下完成。代币计数仅供示例；阈值与 Deep Agents 默认值匹配。","上下文管理只有在代理仍能完成任务并恢复其不再看到的细节时才有用。LangChain 为此保持了针对性的评估：测试在任务中途触发总结，并检查代理是否继续朝着目标前进，以及在大海捞针的案例中，某个事实被总结掉后必须通过文件系统搜索恢复。为了生成足够的事件以比较提示变体，团队在窗口的 10% 到 20% 处触发总结，而非默认的 85%，并在 terminal-bench-2 上使用 Claude Sonnet 4.5 进行了 25% 的触发来研究效果。","在 LangChain 看来，失败的原因是未能关注目标漂移：代理在总结后立即请求澄清，或错误地宣布任务完成。AgentCore 评估提供了一个目标成功率评估器，可以对相同的跟踪进行评分。如果运行测试工具并且未在测试中强制压缩，你尚未知道总结提示会丢失哪些内容。","面向从业者的 AI/ML 新闻和分析，每月有超过 100 万开发者和研究人员阅读。"]},"en":{"title":"Agent Long-Task Context Engineering Analysis: Using Budget Control, Compression, Todo-State, and Memory to Counter Context Overflow and Goal Loss","summary":"The article analyzes four mechanisms in the Agent harness layer for dealing with context overflow and goal loss in long tasks: context budgeting and offloading, compression, todo-state restatement, and cross-session memory.","category":"Industry","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Agent Long-Task Context Engineering Analysis: Using Budget Control, Compression, Todo-State, and Memory to Counter Context Overflow and Goal Loss - Aioga AI News","description":"The article analyzes four mechanisms in the Agent harness layer for dealing with context overflow and goal loss in long tasks: context budgeting and offloading, compression, todo-s...","url":"https://www.aioga.com/en/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:41:19.428Z"},"ja":{"title":"エージェントの長期タスクコンテキストエンジニアリング解析:予算制御、圧縮、TODO状態、メモリを用いてコンテキストオーバーフローやターゲット損失に対抗する","summary":"この記事では、長期間タスクにおけるコンテキストオーバーフローやターゲット損失に対処するためのエージェントハーネス層の4つのメカニズムについて説明します:コンテキスト予算管理とアンロード、圧縮、todo状態の再キャップ、そしてクロスセッションメモリです。","category":"業界動向","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"エージェントの長期タスクコンテキストエンジニアリング解析:予算制御、圧縮、TODO状態、メモリを用いてコンテキストオーバーフローやターゲット損失に対抗する - Aioga AIニュース","description":"この記事では、長期間タスクにおけるコンテキストオーバーフローやターゲット損失に対処するためのエージェントハーネス層の4つのメカニズムについて説明します:コンテキスト予算管理とアンロード、圧縮、todo状態の再キャップ、そしてクロスセッションメモリです。","url":"https://www.aioga.com/ja/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:41:23.678Z"},"ko":{"title":"Agent 장기 과제 컨텍스트 엔지니어링 분석: 예산 관리, 압축, todo-state 및 메모리를 사용하여 컨텍스트 오버플로우와 목표 상실 대응","summary":"논문 분석: Agent harness 계층이 장기 과제에서 컨텍스트 오버플로우와 목표 상실에 대응하는 네 가지 메커니즘: 컨텍스트 예산 및 오프로드, 압축, todo-state 반복, 세션 간 메모리","category":"업계 동향","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Agent 장기 과제 컨텍스트 엔지니어링 분석: 예산 관리, 압축, todo-state 및 메모리를 사용하여 컨텍스트 오버플로우와 목표 상실 대응 - Aioga AI 뉴스","description":"논문 분석: Agent harness 계층이 장기 과제에서 컨텍스트 오버플로우와 목표 상실에 대응하는 네 가지 메커니즘: 컨텍스트 예산 및 오프로드, 압축, todo-state 반복, 세션 간 메모리","url":"https://www.aioga.com/ko/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:41:29.451Z"},"es":{"title":"Análisis de ingeniería contextual de tareas largas con agentes: Utilizar control presupuestario, compresión, estado de tareas y memoria para contrarrestar el desbordamiento de contexto y la pérdida de objetivos","summary":"El artículo explica cuatro tipos de mecanismos en la capa de aprovechamiento de agentes para abordar el desbordamiento de contexto y la pérdida de objetivos en tareas largas: presupuesto y descarga de contexto, compresión, recapitulación de estado pendiente y memoria entre sesiones.","category":"Industria","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Análisis de ingeniería contextual de tareas largas con agentes: Utilizar control presupuestario, compresión, estado de tareas y memoria para contrarrestar el desbordamiento de contexto y la pérdida de objetivos - Aioga Noticias de IA","description":"El artículo explica cuatro tipos de mecanismos en la capa de aprovechamiento de agentes para abordar el desbordamiento de contexto y la pérdida de objetivos en tareas largas: presu...","url":"https://www.aioga.com/es/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:41:32.521Z"},"fr":{"title":"Analyse du projet de contexte long des Agents : Contrôle du budget, compression, état todo et mémoire pour lutter contre le dépassement de contexte et la perte d'objectif","summary":"L'article analyse quatre mécanismes de la couche Agent harness pour faire face aux dépassements de contexte et aux pertes d'objectifs dans les tâches longues : budget et déchargement du contexte, compression, reformulation de l'état todo et mémoire inter-sessions.","category":"Industrie","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Analyse du projet de contexte long des Agents : Contrôle du budget, compression, état todo et mémoire pour lutter contre le dépassement de contexte et la perte d'objectif - Aioga Actualités IA","description":"L'article analyse quatre mécanismes de la couche Agent harness pour faire face aux dépassements de contexte et aux pertes d'objectifs dans les tâches longues : budget et déchargeme...","url":"https://www.aioga.com/fr/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:41:37.653Z"},"de":{"title":"Analyse des Agent-Langzeit-Aufgaben-Kontext: Bekämpfung von Kontextüberlauf und Zielverlust durch Budgetkontrolle, Kompression, Todo-State und Gedächtnis","summary":"Artikelanalyse: Vier Mechanismen der Agent Harness-Ebene zur Bewältigung von Kontextüberlauf und Zielverlust bei Langzeitaufgaben: Kontextbudget und -entlastung, Kompression, Wiederholung des Todo-Status und über Sitzung hinweg erinnerndes Gedächtnis.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Analyse des Agent-Langzeit-Aufgaben-Kontext: Bekämpfung von Kontextüberlauf und Zielverlust durch Budgetkontrolle, Kompression, Todo-State und Gedächtnis - Aioga KI-News","description":"Artikelanalyse: Vier Mechanismen der Agent Harness-Ebene zur Bewältigung von Kontextüberlauf und Zielverlust bei Langzeitaufgaben: Kontextbudget und -entlastung, Kompression, Wiede...","url":"https://www.aioga.com/de/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:41:37.822Z"},"pt-BR":{"title":"Análise da engenharia de contexto de tarefas longas do agente: usando controle de orçamento, compressão, estado-todo e memória para enfrentar o estouro de contexto e a perda de objetivos","summary":"O artigo analisa as quatro categorias de mecanismos da camada Agent harness para lidar com estouro de contexto e perda de objetivos em tarefas longas: orçamento e descarregamento de contexto, compressão, repetição do estado-todo e memória entre sessões.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Análise da engenharia de contexto de tarefas longas do agente: usando controle de orçamento, compressão, estado-todo e memória para enfrentar o estouro de contexto e a perda de objetivos - Aioga Notícias de IA","description":"O artigo analisa as quatro categorias de mecanismos da camada Agent harness para lidar com estouro de contexto e perda de objetivos em tarefas longas: orçamento e descarregamento d...","url":"https://www.aioga.com/pt-BR/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:41:42.903Z"},"ru":{"title":"Анализ инженерии контекста длинных задач агентом: борьба с переполнением контекста и потерей цели с помощью контроля бюджета, сжатия, todo-state и памяти","summary":"Статья анализирует четыре механизма слоя Agent harness для борьбы с переполнением контекста и потерей цели в длинных задачах: бюджет контекста и выгрузка, сжатие, пересказ todo-state и межсессионная память.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Анализ инженерии контекста длинных задач агентом: борьба с переполнением контекста и потерей цели с помощью контроля бюджета, сжатия, todo-state и памяти - Aioga Новости ИИ","description":"Статья анализирует четыре механизма слоя Agent harness для борьбы с переполнением контекста и потерей цели в длинных задачах: бюджет контекста и выгрузка, сжатие, пересказ todo-sta...","url":"https://www.aioga.com/ru/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:41:42.983Z"},"ar":{"title":"تحليل الهندسة السياقية طويلة المهام للوكلاء: استخدم التحكم في الميزانية، الضغط، حالة المهام والذاكرة لمواجهة تجاوز السياق وفقدان الهدف","summary":"تشرح المقالة أربعة أنواع من الآليات في طبقة Agent Harness لمعالجة فائض السياق وفقدان الأهداف في المهام الطويلة: ميزانية السياق وإلغاء التحميل، الضغط، ملخص الحالات المهمة (todo-states)، والذاكرة عبر الجلسات.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"تحليل الهندسة السياقية طويلة المهام للوكلاء: استخدم التحكم في الميزانية، الضغط، حالة المهام والذاكرة لمواجهة تجاوز السياق وفقدان الهدف - Aioga أخبار الذكاء الاصطناعي","description":"تشرح المقالة أربعة أنواع من الآليات في طبقة Agent Harness لمعالجة فائض السياق وفقدان الأهداف في المهام الطويلة: ميزانية السياق وإلغاء التحميل، الضغط، ملخص الحالات المهمة (todo-stat...","url":"https://www.aioga.com/ar/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:41:51.940Z"},"hi":{"title":"एजेंट लम्बे कार्य संदर्भ इंजीनियरिंग विश्लेषण: बजट नियंत्रण, संपीड़न, to-do-स्टेट और मेमोरी का उपयोग करके संदर्भ अधिभार और उद्देश्य हानि से मुकाबला करना","summary":"लेख विश्लेषण करता है कि एजेंट harness स्तर लम्बे कार्यों में संदर्भ अधिभार और उद्देश्य हानि का सामना करने के लिए चार प्रकार की युक्तियाँ अपनाता है: संदर्भ बजट और निष्कासन, संपीड़न, to-do-स्टेट पुनरावृत्ति और सेशन-डाउन मेमोरी।","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"एजेंट लम्बे कार्य संदर्भ इंजीनियरिंग विश्लेषण: बजट नियंत्रण, संपीड़न, to-do-स्टेट और मेमोरी का उपयोग करके संदर्भ अधिभार और उद्देश्य हानि से मुकाबला करना - Aioga AI समाचार","description":"लेख विश्लेषण करता है कि एजेंट harness स्तर लम्बे कार्यों में संदर्भ अधिभार और उद्देश्य हानि का सामना करने के लिए चार प्रकार की युक्तियाँ अपनाता है: संदर्भ बजट और निष्कासन, संपीड़न,...","url":"https://www.aioga.com/hi/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:41:50.034Z"},"it":{"title":"Analisi del contesto di lunga durata degli agenti: controllare il budget, comprimerlo, utilizzare lo stato todo e la memoria per contrastare l'overflow del contesto e la perdita di obiettivi","summary":"L'articolo analizza i quattro meccanismi con cui il livello Agent harness affronta l'overflow del contesto e la perdita di obiettivi nei compiti lunghi: budget del contesto e scarico, compressione, ripetizione dello stato todo e memoria tra sessioni.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Analisi del contesto di lunga durata degli agenti: controllare il budget, comprimerlo, utilizzare lo stato todo e la memoria per contrastare l'overflow del contesto e la perdita di obiettivi - Aioga Notizie IA","description":"L'articolo analizza i quattro meccanismi con cui il livello Agent harness affronta l'overflow del contesto e la perdita di obiettivi nei compiti lunghi: budget del contesto e scari...","url":"https://www.aioga.com/it/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:41:58.045Z"},"nl":{"title":"Analyse van de Agent Lange Taakcontext Engineering: gebruik van budgetcontrole, compressie, todo-state en geheugen om contextoverloop en doelverlies tegen te gaan","summary":"Artikel analyseert de vier mechanismen van het Agent harness-laag om contextoverloop en doelverlies in lange taken aan te pakken: contextbudget en offloaden, compressie, herhaling van todo-state en session-overstijgend geheugen.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Analyse van de Agent Lange Taakcontext Engineering: gebruik van budgetcontrole, compressie, todo-state en geheugen om contextoverloop en doelverlies tegen te gaan - Aioga AI-nieuws","description":"Artikel analyseert de vier mechanismen van het Agent harness-laag om contextoverloop en doelverlies in lange taken aan te pakken: contextbudget en offloaden, compressie, herhaling...","url":"https://www.aioga.com/nl/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:41:58.113Z"},"tr":{"title":"Ajan uzun görevli bağlamsal mühendislik analizi: Bağlam taşmasını ve hedef kaybını önlemek için bütçe kontrolü, sıkıştırma, yapılacak durum ve bellek kullanın","summary":"Makale, uzun görevlerde bağlam taşması ve hedef kaybını ele almak için Ajan harness katmanında dört tür mekanizmayı açıklar: bağlam bütçeleme ve boşaltma, sıkıştırma, yapılacak durum özetleri ve oturumlar arası bellek.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Ajan uzun görevli bağlamsal mühendislik analizi: Bağlam taşmasını ve hedef kaybını önlemek için bütçe kontrolü, sıkıştırma, yapılacak durum ve bellek kullanın - Aioga AI Haberleri","description":"Makale, uzun görevlerde bağlam taşması ve hedef kaybını ele almak için Ajan harness katmanında dört tür mekanizmayı açıklar: bağlam bütçeleme ve boşaltma, sıkıştırma, yapılacak dur...","url":"https://www.aioga.com/tr/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:42:06.853Z"},"vi":{"title":"Phân tích kỹ thuật bối cảnh nhiệm vụ dài của Agent: Sử dụng kiểm soát ngân sách, nén, trạng thái todo và bộ nhớ để chống tràn bối cảnh và mất mục tiêu","summary":"Bài viết phân tích bốn cơ chế của lớp Agent harness để đối phó với tràn bối cảnh và mất mục tiêu trong nhiệm vụ dài: ngân sách bối cảnh và chuyển tải, nén, nhắc lại trạng thái todo và bộ nhớ xuyên phiên.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Phân tích kỹ thuật bối cảnh nhiệm vụ dài của Agent: Sử dụng kiểm soát ngân sách, nén, trạng thái todo và bộ nhớ để chống tràn bối cảnh và mất mục tiêu - Tin tức AI Aioga","description":"Bài viết phân tích bốn cơ chế của lớp Agent harness để đối phó với tràn bối cảnh và mất mục tiêu trong nhiệm vụ dài: ngân sách bối cảnh và chuyển tải, nén, nhắc lại trạng thái todo...","url":"https://www.aioga.com/vi/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:42:03.291Z"},"id":{"title":"Analisis Proyek Konteks Tugas Panjang Agent: Menggunakan Kontrol Anggaran, Kompresi, todo-state, dan Memori untuk Melawan Overflow Konteks dan Kehilangan Tujuan","summary":"Artikel ini menganalisis empat mekanisme lapisan Agent harness untuk menangani overflow konteks dan kehilangan tujuan dalam tugas panjang: anggaran konteks dan offload, kompresi, pengulangan todo-state, dan memori lintas sesi.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Analisis Proyek Konteks Tugas Panjang Agent: Menggunakan Kontrol Anggaran, Kompresi, todo-state, dan Memori untuk Melawan Overflow Konteks dan Kehilangan Tujuan - Berita AI Aioga","description":"Artikel ini menganalisis empat mekanisme lapisan Agent harness untuk menangani overflow konteks dan kehilangan tujuan dalam tugas panjang: anggaran konteks dan offload, kompresi, p...","url":"https://www.aioga.com/id/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:42:12.611Z"},"th":{"title":"การวิเคราะห์วิศวกรรมบริบทงานยาวของ Agent: ใช้งบประมาณควบคุม, การบีบอัด, todo-state และความทรงจำต่อสู้กับการล้นบริบทและการสูญเสียเป้าหมาย","summary":"บทความวิเคราะห์กลไกสี่ประเภทของ Agent harness ชั้นสำหรับจัดการการล้นบริบทและการสูญเสียเป้าหมายในงานยาว: งบประมาณบริบทและการระบาย, การบีบอัด, การทวนซ้ำ todo-state และความทรงจำข้ามเซสชัน","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"การวิเคราะห์วิศวกรรมบริบทงานยาวของ Agent: ใช้งบประมาณควบคุม, การบีบอัด, todo-state และความทรงจำต่อสู้กับการล้นบริบทและการสูญเสียเป้าหมาย - ข่าว AI Aioga","description":"บทความวิเคราะห์กลไกสี่ประเภทของ Agent harness ชั้นสำหรับจัดการการล้นบริบทและการสูญเสียเป้าหมายในงานยาว: งบประมาณบริบทและการระบาย, การบีบอัด, การทวนซ้ำ todo-state และความทรงจำข้ามเซ...","url":"https://www.aioga.com/th/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:42:14.514Z"},"pl":{"title":"Analiza inżynierii kontekstu dla długoterminowych zadań Agenta: kontrola budżetu, kompresja, stan-to-do i pamięć w przeciwdziałaniu przepełnieniu kontekstu i utracie celów","summary":"Artykuł analizuje cztery mechanizmy warstwy Agent harness w radzeniu sobie z przepełnieniem kontekstu i utratą celów podczas długoterminowych zadań: budżet i odciążanie kontekstu, kompresja, powtarzanie stanu-to-do i pamięć między sesjami.","category":"行业动态","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Analiza inżynierii kontekstu dla długoterminowych zadań Agenta: kontrola budżetu, kompresja, stan-to-do i pamięć w przeciwdziałaniu przepełnieniu kontekstu i utracie celów - Aioga Wiadomości AI","description":"Artykuł analizuje cztery mechanizmy warstwy Agent harness w radzeniu sobie z przepełnieniem kontekstu i utratą celów podczas długoterminowych zadań: budżet i odciążanie kontekstu,...","url":"https://www.aioga.com/pl/news/cmtzezuk0077wrowyi7a4qy7v/","contentTranslated":true,"sourceHash":"fcf1ecbbd461fe0a","translatedAt":"2026-09-13T06:42:20.767Z"}},"evidenceTier":"verified-news","reviewStatus":"editorial-selected","indexable":true,"editorialCover":""}}