One dial for how much work gets handed off — the main model keeps the judgment, helpers do the rest一个档位决定多少工作被转交——主模型保留判断力，其余交给帮手
Frontier models are excellent planners and expensive typists. The cheapest way to run an agent is to stop paying flagship prices for the parts that do not need flagship judgment.顶级模型是优秀的规划者，也是昂贵的打字员。运行 Agent 最省钱的方式，就是不再为不需要顶级判断力的部分支付旗舰价格。
Any subtask that is mechanical and can be fully specified in writing goes to a cheap executor. On rote work the measured saving is roughly 2.3–2.7× versus running the same step inline on the main model — and for repetitive, template-shaped work the small model is often more accurate, because it follows the written contract instead of improvising.任何机械性、能用书面形式完整规格化的 subtask 都交给便宜的执行模型。在机械性工作上，相比在主模型上内联执行同一步骤，实测节省约 2.3–2.7×——而且对重复性、模板化的活儿，小模型常常更准，因为它照书面契约执行，而不是即兴发挥。
Judgment that cannot be specified goes the other way: up. A stronger advisor reviews the hardest calls. The main model keeps the three things it is genuinely needed for — planning, acceptance checks, and synthesis of results into the final answer.无法规格化的判断则走另一条路：向上。更强的 advisor 复核最难的判断。主模型保留三件真正需要它的事——规划、验收检查、把结果综合成最终答复。
Smart is deliberately one dial, not a settings page. It controls who does the grunt work — never which model answers you.Smart 刻意做成一个档位，而不是一页设置。它控制的是谁干粗活——而不是谁来回答你。
| Level档位 | Posture姿态 | What changes具体变化 |
|---|---|---|
| off | Solo单干 | Everything runs inline on the main model. No helpers, no extra cost, no oversight.一切在主模型上内联执行。没有帮手，没有额外成本，也没有把关。 |
| on | Balanced均衡 | Delegation for clearly mechanical work; the advisor is available when the agent is stuck or a claim needs an independent check.明显机械性的工作走委派；Agent 卡住或结论需要独立核验时可请 advisor。 |
| anchor | Cost-first成本优先 | Aggressive delegation: mechanical work is handed off by default and the main model keeps its turns short and dense. The advisor is off for strong models — but stays available as a quality floor when a budget-tier model is answering (e.g. routed there by Conductor Mode), because on those turns delegation is gated off and the advisor is the only helper left.激进委派：机械性工作默认转交，主模型保持回合短而密。强模型轮次 advisor 关闭——但当回答者是廉价档模型时（例如被 Conductor Mode 路由过去），advisor 保留为质量地板：那些轮次委派被档位闸门挡掉，advisor 是仅剩的帮手。 |
| guard | Quality-first质量优先 | Delegation as usual, plus advisor verification before any non-trivial deliverable is declared done.委派照常，另加一道 advisor 裁决：任何非平凡交付在宣布完成前必须过审。 |
anchor and guard are the two opinionated ends: one optimizes spend, the other optimizes trust.anchor 与 guard 是两个立场鲜明的端点：一个优化开销，一个优化可信度。
The design rule for the user surface: no manual configuration, ever. You state intent; the agent reconfigures itself.用户界面的设计铁律：永远不需要手动配置。你说出意图，Agent 自己完成配置。
/smart shows the current level, the executor chain, the advisor model, and per-turn budgets in one panel; /smart anchor flips the dial. Everything finer-grained is conversational: say "use a stronger executor", "raise the delegation budget to $2", or "switch the advisor to Opus" — the agent applies the change through its own config tool and confirms what it changed. There is no settings file the user is ever pointed at./smart 一屏展示当前档位、执行链、advisor 模型与单轮预算；/smart anchor 直接换档。更细粒度的一切都用对话完成：说"换个更强的 executor"、"委派预算提到 $2"、"advisor 换成 Opus"——Agent 通过自己的配置工具落实变更并复述改了什么。用户永远不会被指向某个配置文件。
Defaults are chosen so the dial works out of the box: a cheap executor with a stronger fallback behind it (so a flaky cheap model degrades instead of failing the delegation), and a reasoning-effort suffix (model@xhigh) for the rare delegation that needs deeper thinking at executor prices.默认值保证开箱即用：便宜的 executor 背后挂一个更强的兜底（便宜模型抽风时是降级而不是整体失败），并支持推理力度后缀（model@xhigh）——少数需要深思考的委派可以用 executor 的价格买到更强推理。
A delegation is not a hint or a suggestion. It is a literal contract, written before the work starts.一次委派不是提示，也不是建议，而是一份字面意义上的契约，在工作开始之前就写好。
The contract states the GOAL, the INPUTS (explicit file paths — the executor sees nothing else, not the conversation), the OUTPUT format, and the forbidden side effects. Acceptance criteria are written before dispatch and are machine-checkable: the parent runs them itself on the executor's output, instead of trusting the executor's own report. The executor runs in an isolated mini-session with file tools and bash, so a bad delegation cannot disturb the parent's state.契约写明 GOAL（目标）、INPUTS（输入：明确的文件路径——执行者看不到其他任何东西，包括对话本身）、OUTPUT 格式，以及禁止的副作用。验收标准在派发之前写好，并且可被机器检查：由父方自己对着执行者的产出运行这些标准，而不是相信执行者自己的报告。执行者在带文件工具和 bash 的隔离迷你会话中运行，一次糟糕的委派也不会打扰父方的状态。
Delegations run in two shapes. Synchronous is the default: the parent blocks, gets the finished work back, verifies it, moves on. Background (background=true) returns a delegation_id immediately; the parent keeps working and collects the result once, in a later turn, via check_id. Background runs are capacity-capped per user with a TTL, and a rejected start never consumes the turn's delegation quota — validation happens before the slot is reserved.委派有两种形态。同步是默认：父方阻塞等待，拿回成品、验收、继续。后台（background=true）立即返回 delegation_id；父方继续干活，在之后的回合用 check_id 一次性收取结果。后台运行按用户设并发上限和 TTL，且被拒绝的启动永远不消耗本轮委派配额——校验先于占位发生。
Coordination is not free: the parent pays for every handoff, so tiny tasks stay inline, and one batched spec beats many small calls.协调并非免费：父方为每一次交接付费，因此微小任务留在内联执行，而一份批量规格胜过许多次小调用。
Executors run on budgets — a timeout, an iteration cap, a cost cap. Hitting one used to mean the whole delegation "failed" and the parent redid everything. That threw away real work.执行者跑在预算之内——超时、轮数上限、成本上限。撞上任何一个，过去都意味着整次委派"失败"，父方从头重做。这等于把真实完成的工作扔掉。
Now a budget stop triggers salvage: the run's transcript is folded into a DONE / PENDING hand-off and returned as a partial success. The parent is told exactly what was completed and continues from PENDING instead of restarting. Two honesty rules keep this from becoming a lie. First, partial requires evidence: at least one tool action must have actually succeeded — a run whose every call failed, or that only produced "working on it…" prose, returns a plain failure. Second, models that emit no prose at all (pure tool-callers often don't) get their hand-off synthesized from the action log — evidence over prose, so completed work is never invisible just because nobody narrated it.现在预算截停触发打捞（salvage）：运行轨迹被折叠成一份 DONE / PENDING 交接单，作为 partial 成功返回。父方确切知道哪些已完成，从 PENDING 续做而不是重启。两条诚实规则防止它变成谎言。第一，partial 需要证据：至少一条工具动作真正成功过——所有调用全失败、或只产出"正在处理中…"式空话的运行，照旧返回普通失败。第二，全程不写文字的模型（纯工具调用型模型常常如此）的交接单从动作日志合成——证据优先于文字，已完成的工作不会因为没人叙述就变得不可见。
The same honesty extends to the books: partial is its own outcome in stats and telemetry — not a success that hides problems, not a failure that understates the executor. And long runs protect themselves: the executor prunes its own context mid-run, spilling full tool outputs to disk (size-capped, swept after seven days, deleted when a run finishes cleanly) so a delegation dies of budget, not of context bloat.同样的诚实延伸到账本：partial 在统计与遥测中是独立的一类结果——既不是掩盖问题的成功，也不是低估执行者的失败。长任务还会自我保护：executor 在运行中途裁剪自己的上下文，把完整工具输出落盘（限大小、七天清扫、干净跑完即自删），让委派死于预算，而不是死于上下文膨胀。
The advisor is one system with three depths of intervention — advise, judge, drive — and the differences between them are deliberately stark.Advisor 是一个系统的三种介入深度——动嘴、盖章、上手——三者之间的差异刻意做得非常分明。
Consult sends a recent slice of the conversation to an advisor and asks for a plan or a correction when the agent is genuinely stuck. Verify is the opposite: an independent judge with no access to the conversation sees only the claim being made plus the evidence pasted next to it, and returns a verdict of PASS or FAIL — used before declaring any non-trivial deliverable done. Takeover is the deepest and never asked for: when a budget-tier executor keeps failing the same way or loops, a strong model is pinned as the executor for a turn or two, does the work directly, and hands back a written summary. Rescue always goes upward — a strong model is never taken over, and the rescuer is never a cheap one.咨询（consult）把最近一段对话交给 advisor，在 Agent 真正卡住时请它给方案或纠正。裁决（verify）恰恰相反：一位没有对话访问权的独立评审，只看被声称的结论以及旁边粘贴的证据，给出 PASS 或 FAIL 的判定——用于宣布任何非平凡交付完成之前。接管（takeover）介入最深，且从不由 Agent 主动请求：当廉价档执行者反复以同样方式失败或陷入循环，一个强模型会被钉为执行者开车一到两轮，直接完成工作，再交回一份书面交接。救援永远向上——强模型不会被接管，救援者也绝不是廉价模型。
The no-context design of verify is the point: a judge who saw the agent's reasoning would inherit the agent's own biases and simply agree. A judge who sees only the claim and the evidence has no choice but to check the evidence.裁决不看上下文正是关键：看过 Agent 推理过程的评审会继承 Agent 自己的偏见，然后直接同意。只看结论和证据的评审别无选择，只能去核对证据。
Cost discipline follows the same shape. Consults are frequent and carry a context snapshot, so they run on a fast cheap model and the snapshot is held under a hard size ceiling; verify and takeover are rare, bounded calls where judgment is the product, so they stay on a strong model. Access is earned, not timed: instead of waiting a fixed number of turns, the agent may consult early by showing its work — a real attempt with the actual error attached — while restated questions and filler are turned away. And all three forms spend from one shared budget: per-turn and per-session use caps plus a session dollar ceiling, measured from real provider costs.成本纪律沿同一形状展开。咨询高频且携带上下文快照，因此跑在快而便宜的模型上，快照有硬性体积上限；裁决与接管低频且有界，判断力本身就是产品，因此保留强模型。资格靠证据挣得，而非按时间等待：Agent 不必干等固定轮数，只要出示工作过程——附上真实报错的一次实际尝试——就能提前咨询；而复述问题与凑字数会被拒之门外。三种形态花的是同一份预算：每轮与每会话的次数上限，加上按供应商真实成本累计的会话美元硬顶。
Knowing delegation is cheaper does not mean the model reaches for it. Mid-task, the switching cost always feels bigger than the remaining work. Anchor counters that bias with three layers of soft guidance — and deliberately stops short of enforcement.知道委派更便宜，不等于模型会伸手去用。任务进行中，切换成本永远显得比剩余工作量大。anchor 档用三层软引导对抗这种偏差——并刻意止步于强制。
Layer one is a tripwire rule in the system prompt: a third consecutive similar mechanical call means a delegation was already missed — stop, batch the remainder into one spec. Layer two puts anchor hints in the tool descriptions of exactly the tools where the mistake happens (bash, file edits, file writes), so the reminder is adjacent to the temptation. Layer three is a one-shot checkpoint: when the runtime detects a streak of mechanical calls mid-run, it injects a single system note asking the model to reconsider — once per run, stripped before the next run begins.第一层是系统提示里的绊线规则：连续第三次同类机械调用，意味着委派已经被错过——停下，把剩余工作打包成一份规格。第二层把锚点提示写进工具描述，且恰好写在错误发生的那几个工具上（bash、文件编辑、文件写入），让提醒紧贴诱惑。第三层是一次性检查点：运行时在途中检测到连续机械调用时，注入一条系统提示请模型重新考虑——每次运行只注入一次，下次运行开始前即被剥离。
Two constraints shaped all three layers. Everything is appended at the context tail and ephemeral, so the prompt cache prefix is never invalidated — guidance must not cost more than it saves. And every layer carries an escape clause: work that genuinely needs live conversation context stays inline, no questions asked. Hard interception was considered and rejected — a model forced to delegate writes worse specs than a model persuaded to.三层共同受两条约束塑形。所有注入都追加在上下文尾部且短暂存在，prompt cache 前缀永不失效——引导的成本不能超过它省下的钱。每一层都保留逃生条款：真正需要实时对话上下文的工作留在内联，无需解释。硬拦截被认真考虑过，然后被否决——被强迫委派的模型写出的规格，比被说服委派的模型更差。
Conductor Mode and Smart solve different problems, so they compose.Conductor Mode 与 Smart 解决的是不同问题，因此可以组合使用。
Conductor picks the model for each user turn, deciding which lane a conversation travels in. Smart reshapes cost inside a single turn: the main model buys a cheap execution and — on guard — an expensive review, without leaving the conversation. Everything underneath is configurable per agent (the executor chain, the advisor model, the cost caps), but by design the agent does that configuring itself, in conversation, on request.Conductor 为每一轮用户回合选模型，决定一段对话走哪条车道。Smart 在单轮内部重塑成本：主模型买到一次便宜的执行，在 guard 档还买到一次昂贵的复核，全程无需离开会话。底层一切都按 agent 可配置（执行链、advisor 模型、成本上限），但按设计，这些配置由 Agent 在对话中按需自己完成。