7 天 Agent Harness:让 Agent 无人值守跑一整夜
课程题库 · 28 / 28 道题
标签
全部#long-horizon6#state-management6#acceptance-criteria2#assertions2#capability-boundary2#git2#initialization2#mutation-testing2#persistence2#rollback2#verification2#architecture1
还有 53 个标签收起标签
#benchmarks1#boundaries1#budget1#checkpoint1#circuit-breaker1#compaction1#constraint-fidelity1#constraint-pinning1#constraints1#context-engineering1#context-window1#cost1#course-recap1#crash-resume1#crash-safety1#developer-experience1#end-to-end1#evaluation1#experiment-design1#failure-handling1#failure-modes1#false-completion1#false-green1#false-progress1#feature-list1#gates1#governance-decay1#green-point1#harness-design1#harness-engineering1#human-readable1#idempotence1#interface-boundary1#intervention1#model-capability1#module-merge1#probe-blind-spot1#probe-design1#progress-file1#real-time-control1#report-design1#reporting1#risk-exposure1#safety-invariants1#self-verification1#silent-failure1#single-source1#stall-detection1#stop-reason1#switch-experiment1#trade-offs1#tradeoff-reasoning1#trust-vs-throughput1
D1 为什么 Agent 跑过第二个上下文窗口就开始退化
上下文窗口和状态有什么区别?为什么长时程 Agent 的权威进度不能只存在窗口里?What is the difference between the context window and state? Why can a long-horizon agent's authoritative progress not live only in the window?
国内高频海外高频进阶#long-horizon#state-management#context-window🔒 解锁 D1 后查看答题要点。
为什么给模型的接口不该知道当前开了哪些 harness 功能?如果知道了会发生什么?Why should the model-facing interface not know which harness features are switched on? What happens if it does?
国内高频海外高频深入#experiment-design#interface-boundary#false-green🔒 解锁 D1 后查看答题要点。
一个无人值守跑了一夜的 Agent 早上告诉你全部任务已完成。你会怎么验证这句话?An agent that ran unattended overnight tells you in the morning that every task is complete. How do you verify that claim?
国内高频海外高频进阶#verification#long-horizon#reporting🔒 解锁 D1 后查看答题要点。
同一个模型、同一套提示词,换一个 harness 成绩提升十几个点。这说明了什么?Same model, same prompts, but swapping the harness lifts the benchmark score by more than ten points. What does that tell you?
国内高频海外高频进阶#harness-engineering#benchmarks#model-capability🔒 解锁 D1 后查看答题要点。
D2 状态与上下文的分界:把权威状态搬出窗口
设计一个长时程 Agent 的状态层,你会把哪些东西写进磁盘,哪些坚决不写?Designing the state layer for a long-horizon agent, what goes to disk and what stays out of it?
国内高频海外高频进阶#state-management#long-horizon#persistence🔒 解锁 D2 后查看答题要点。
窗口重置后,把之前的完整对话历史重新喂给模型是个好主意吗?After a window reset, is replaying the full prior transcript into the model a good idea?
国内高频海外高频进阶#context-engineering#long-horizon#cost🔒 解锁 D2 后查看答题要点。
你加了一个状态层,跑下来指标变好了。怎么证明是它起的作用,而不是碰巧?You added a state layer and the metrics improved. How do you show the state layer caused it, rather than coincidence?
国内高频海外高频深入#evaluation#mutation-testing#state-management🔒 解锁 D2 后查看答题要点。
状态文件在写入过程中进程被杀了怎么办?What happens if the process is killed while the state file is being written?
国内高频海外高频深入#crash-safety#persistence#state-management🔒 解锁 D2 后查看答题要点。
D3 initializer agent:init.sh、进度文件与第一个 commit
让一个 Agent 无人值守地开发一个项目,开工前你会先让它做哪几件事?Before letting an agent develop a project unattended overnight, what do you have it do first?
国内高频海外高频进阶#initialization#failure-modes#long-horizon🔒 解锁 D3 后查看答题要点。
进度文件应该写什么?写成给人看的日志和给下一个窗口看的摘要有什么区别?What belongs in a progress file, and how does a human-readable log differ from a summary written for the next window?
国内高频海外高频进阶#progress-file#state-management#long-horizon🔒 解锁 D3 后查看答题要点。
为什么初始化要独立成一个阶段,而不是让主循环第一轮顺手做掉?Why make initialization its own phase instead of letting the first pass of the main loop handle it?
国内高频海外高频深入#architecture#initialization#failure-handling🔒 解锁 D3 后查看答题要点。
怎么判断一个工作区对新来的 Agent 足够友好?给一个可执行的检验方法。How do you judge whether a workspace is friendly enough for a newly arrived agent? Give an executable test.
国内高频海外高频深入#acceptance-criteria#verification#developer-experience🔒 解锁 D3 后查看答题要点。
D4 一次只做一件事:feature 清单与 git 纪律
什么情况下 Agent 可以自己执行 git commit,什么情况下绝对不行?When may an agent run git commit on its own, and when must it never?
国内高频海外高频进阶#git#state-management#boundaries🔒 解锁 D4 后查看答题要点。
怎么强制一个 Agent 一次只做一件事?写在提示词里够吗?How do you force an agent to do one thing at a time? Is putting it in the prompt enough?
国内高频海外高频进阶#feature-list#constraints#harness-design🔒 解锁 D4 后查看答题要点。
Agent 做坏了一条 feature,你会让它回滚还是继续修?依据是什么?An agent botches one feature - do you roll back or let it keep fixing? On what basis?
国内高频海外高频深入#rollback#green-point#risk-exposure🔒 解锁 D4 后查看答题要点。
让 Agent 操作 git 有哪些具体的危险?你会怎么在代码层面防住?What can go wrong when an agent drives git, and how do you prevent it in code?
国内高频海外高频深入#git#safety-invariants#assertions🔒 解锁 D4 后查看答题要点。
D5 自验证:端到端闸门、谎报完成与打转的自动干预
Agent 报告任务完成,你的 harness 该用什么依据决定信不信?An agent reports a task done - what evidence should your harness use to decide whether to believe it?
国内高频海外高频进阶#self-verification#false-completion#gates🔒 解锁 D5 后查看答题要点。
端到端验证用浏览器自动化和只打 HTTP 接口,各有什么得失?Browser automation versus HTTP-only end-to-end verification - what does each buy and cost?
国内高频海外高频进阶#end-to-end#trade-offs#capability-boundary🔒 解锁 D5 后查看答题要点。
你怎么发现一个 Agent 在原地打转?发现之后该做什么?How do you detect an agent spinning in place, and what do you do once you have?
国内高频海外高频深入#stall-detection#intervention#rollback🔒 解锁 D5 后查看答题要点。
一个防谎报的验证器,你怎么确认它自己不是恒真的?A verifier built to catch false completions - how do you confirm it is not tautological itself?
国内高频海外高频深入#assertions#mutation-testing#probe-design🔒 解锁 D5 后查看答题要点。
D6 崩溃续跑、预算熔断与 governance decay
一个跑了六小时的 Agent 进程被杀了。你希望它下次启动时怎么恢复?An agent process that had been running for six hours gets killed - how should it come back on the next start?
国内高频海外高频深入#crash-resume#checkpoint#idempotence🔒 解锁 D6 后查看答题要点。
预算熔断和事后成本报表是两件事吗?分别解决什么问题?Are a budget circuit breaker and a post-hoc cost report two different things, and what does each solve?
国内高频海外高频进阶#budget#circuit-breaker#real-time-control🔒 解锁 D6 后查看答题要点。
反复压缩上下文会让 Agent 逐渐不遵守最初的安全约束。你会怎么防?Repeated context compaction makes an agent drift away from its original safety constraints - how would you defend against that?
国内高频海外高频深入#governance-decay#constraint-pinning#compaction🔒 解锁 D6 后查看答题要点。
你已经有一套压缩后的事实校验探针了,为什么它发现不了约束被擦掉?You already have probes that check factual fidelity after compaction - why do they miss constraints being erased?
国内高频海外高频深入#probe-blind-spot#constraint-fidelity#capability-boundary🔒 解锁 D6 后查看答题要点。
D7 综合:让它跑一整夜,早上验收
六个独立开发的模块合并进一个系统,你最担心出现什么问题?怎么发现它?Six modules developed independently get merged into one system - what worries you most, and how would you find it?
国内高频海外高频深入#silent-failure#module-merge#switch-experiment🔒 解锁 D7 后查看答题要点。
设计一份无人值守运行的早间报告,你会放哪些内容、砍掉哪些内容?Design the morning report for an unattended run - what goes in, and what gets cut?
国内高频海外高频进阶#report-design#human-readable#single-source🔒 解锁 D7 后查看答题要点。
Agent 跑了一夜完成了三十条里的十八条。这个结果算好还是不好?An agent ran overnight and finished 18 of 30 items - is that a good result or a bad one?
国内高频海外高频深入#acceptance-criteria#false-progress#stop-reason🔒 解锁 D7 后查看答题要点。
回顾整套 harness,如果只能保留一个模块,你留哪个?为什么?Looking back at the whole harness, if you could keep only one module, which one and why?
国内高频海外高频深入#tradeoff-reasoning#trust-vs-throughput#course-recap🔒 解锁 D7 后查看答题要点。