01 · Agent infrastructure, late 202601 · 2026 年底的代理基础设施
Everything agents need
is being rebuilt.代理要用的每一样东西，
都在被重做一遍。
Every direction we found in YC's Summer 2026 batch, YC's fall requests and a16z's 2026 bets: what the companies do, how we'd start, where the work is, and who pays.YC 夏季批次、YC 秋季创业需求和 a16z 今年的投资里，我们找到的所有方向：公司在做什么、我们怎么起步、力气花在哪、谁付钱。
235
companies in YC's Summer 2026 batch家公司在 YC 2026 夏季批次
~30
of them build infrastructure for agents (our count)家在给代理做基础设施（按简介数）
8
a16z deals from 2026, all listed below笔 a16z 今年的投资，下面都列了
17
directions, one row each个方向，每个一行
Weight: where most of the hard work goes偏重：最费力气的地方
Engineering工程
Product产品
Sales销售
Direction · weight方向 · 偏重What the companies do公司在做什么How we'd start我们怎么起步Who pays谁付钱
A
Running agents让代理跑起来
01Agent computers代理的电脑
Engineering工程
What the companies do公司在做什么
- machine0YC S26
Agents rent their own cloud VMs from a CLI: snapshot, clone, pay by the minute.代理用命令行自己开云主机，能快照、克隆，按分钟付费。
- Runtaa16z seed
A full, stateful computer for agents, local or cloud, with enterprise policy. “Not yet another sandbox cloud.”给代理一台完整、有状态的电脑，本地云上都能跑，带企业级权限。a16z 原话：“不是又一个沙盒云”。
How we'd start我们怎么起步One primitive agents can rent themselves, over CLI or MCP.做一个代理能自己租用的基础件，走命令行或 MCP。
Who pays谁付钱Agent builders, by usage; enterprises later. Crowded: E2B, Daytona, Modal, the big clouds.做代理的团队按用量付费，企业在后面。已经很挤：E2B、Daytona、Modal 和各大云厂商。
02Durable execution持久执行
Engineering工程
What the companies do公司在做什么
- Temporala16z Series D
Long agent jobs survive crashes and retries. OpenAI and Replit run on it.长时间的代理任务崩了、重试了也能接着跑，OpenAI 和 Replit 在用。
- a16zBig Ideas 2026
One agent goal fans out into thousands of subtasks, so routing, locking, state and policy all need rebuilding.一个代理目标能扇出成几千个子任务，路由、加锁、状态、策略都得重做。
How we'd start我们怎么起步An open-source library that drops into existing agent frameworks.做一个能直接塞进现有代理框架的开源库。
Who pays谁付钱Platform teams at AI-native companies. Temporal is already the default.AI 原生公司的平台团队。Temporal 已经是默认选择。
03Inference cost推理成本
Engineering工程
What the companies do公司在做什么
- UnderstudyYC S26
Trains a small open model on your production traffic, checks it with evals, then routes routine calls to it.用你的线上流量训一个小开源模型，用评测把关，再把日常请求转给它。
- Riften · ConiferYC S26
Gateways that send each request to the cheapest model that can handle it, in the cloud or on the device.网关：把每个请求转给能胜任的最便宜的模型，云端或本地。
- Gimleta16z Growth
Pools GPUs, CPUs and accelerators into one inference resource.把 GPU、CPU 和加速卡合成一个推理资源池。
How we'd start我们怎么起步A drop-in gateway: change one base URL, take a share of the savings.做一个即插即用的网关：改一个地址就能用，从省下的钱里分成。
Who pays谁付钱Teams with big model bills. Savings are easy to prove; falling model prices are the risk.模型账单大的团队。省了多少一算就清楚；风险是模型本身在降价。
04Observability可观测
Product产品
What the companies do公司在做什么
- HyperprobeYC S26
Debugger-level visibility into agents running live.对线上运行的代理做调试器级别的观察。
- Agnost AI · BuildboxYC S26
Find failure patterns in production conversations; tie agent runs to user outcomes.在线上对话里找失败模式；把代理的运行和用户结果对上。
- CodagYC S26
Compresses logs for AI incident-response agents.给处理线上事故的 AI 代理压缩日志。
How we'd start我们怎么起步An SDK that shows why an agent failed in production, free to start.做一个能看出代理为什么在线上出错的 SDK，免费起步。
Who pays谁付钱Agent teams already in production. Incumbents: LangSmith, Braintrust, Arize.已经上线的代理团队。老玩家：LangSmith、Braintrust、Arize。
B
Trusting agents让代理可信
05Identity身份
Product产品Sales销售
What the companies do公司在做什么
- InkboxYC S26
One API gives an agent its own email, phone number, 2FA vault and iMessage.一个 API 给代理发邮箱、手机号、二次验证保险箱和 iMessage。
- Keycarda16z
Credentials tied to one agent and one task, revocable in real time.绑定到某个代理、某项任务的凭证，可以实时吊销。
- a16zBig Ideas 2026
“Know your agent”: signed proof of who an agent acts for and what it may do.“认识你的代理”（KYA）：用签名凭证写明代理替谁办事、能做什么。
How we'd start我们怎么起步A developer API first; enterprise policy later.先做开发者 API，再往企业权限管理走。
Who pays谁付钱Agent builders self-serve; enterprises pay the big contracts.做代理的团队自助接入；大单来自企业。
06Security and control安全与管控
Sales销售
What the companies do公司在做什么
- Neoa16z seed
Endpoint security that can tell the person from their agent.分得清是人还是代理的终端安全。
- doxxneta16z Series A
A private network for people and their agents.给人和他们的代理用的私有网络。
- Armadina16z Series B
A team of AI agents that attacks you the way hackers would.用一群 AI 代理像黑客那样攻击你，找出漏洞。
- OneCLI · Bylaw · DecaworkYC S26
Keep keys away from agents; block rule-breaking actions; a control panel for IT.不让代理碰密钥；运行时拦下违规动作；给 IT 团队的代理管控台。
How we'd start我们怎么起步One control a security team can switch on in a day.做一个安全团队一天就能打开的管控功能。
Who pays谁付钱Security and IT leaders. Big contracts, long cycles, SOC 2 first.企业的安全和 IT 负责人。单子大、周期长，先过 SOC 2。
What the companies do公司在做什么
- AgentcardYC S26
Single-use debit cards, so an agent can pay without your real card.给代理发一次性借记卡，不用交出你的真卡。
- RindlerYC S26
Commerce infrastructure for agents.代理购物用的基础设施。
- YC requestNemil Dalal
“Agents are going to use crypto networks as financial rails.”“代理会把加密网络当作金融轨道。”
How we'd start我们怎么起步One purchase flow, with a card-issuing or stablecoin partner.找一家发卡方或稳定币合作方，先打通一种购买场景。
Who pays谁付钱Apps whose agents buy for users. Regulated; partners come first.替用户下单的代理应用。受监管，先谈合作方。
C
Making agents useful让代理有用
What the companies do公司在做什么
- GlenYC S26
Memory shared across agent sessions and work tools.在代理会话和办公工具之间共享的记忆。
- screenpipeYC S26
Memory built locally from what’s on your screen.在本地从屏幕内容里建起来的记忆。
- Egoist MachinesYC S26
Personal memory you carry between AI apps.在不同 AI 应用之间随身带走的个人记忆。
How we'd start我们怎么起步An open-source SDK that plugs into popular agents.做一个能接进主流代理的开源 SDK。
Who pays谁付钱Developers, freemium. Risk: model vendors build it in.开发者，免费加增值。风险：模型厂商自己做了。
09Tools and data工具与数据
Product产品
What the companies do公司在做什么
- Exaa16z
Search built for agents; Cursor and Cognition use it.专给代理用的搜索，Cursor 和 Cognition 在用。
- Context · ClickYC S26
Structured web data; context beyond web search, over MCP.结构化网页数据；通过 MCP 提供网页搜索之外的信息。
- ExecutorYC S26
An open-source MCP gateway.开源的 MCP 网关。
How we'd start我们怎么起步One tool agents can find and pay for themselves.做一个代理能自己找到、自己付费的工具。
Who pays谁付钱Agent builders, by usage. Agents themselves are the sales channel.做代理的团队，按用量付费。获客渠道就是代理自己。
10Software built for agents为代理而建的软件
Product产品Sales销售
What the companies do公司在做什么
- a16zMay 2026 essay
“An agent doesn’t need a browser. It needs an API, context, instructions, and the ability to act.”“代理不需要浏览器，它需要 API、上下文、指令和动手的权限。”
- a16zBig Ideas 2026
Old systems of record lose ground. “We’re no longer designing for humans, but for agents.”老的记录系统在失势。“我们不再为人设计，而是为代理设计。”
How we'd start我们怎么起步Rebuild one industry’s system of record agent-first: API, permissions, audit, rollback.挑一个行业，把它的核心记录系统按代理优先重做：API、权限、审计、回滚。
Who pays谁付钱Mid-size firms in that industry. Long sales, very sticky.这个行业里的中型企业。销售周期长，但用上就很难换。
D
Proving agents did it right证明代理做对了
11Test sandboxes and evals测试沙盒与评测
Engineering工程Product产品
What the companies do公司在做什么
- ArchalYC S26
Sandboxed Slack, Linear and Datadog APIs for agents; reset in milliseconds.给代理用的沙盒版 Slack、Linear、Datadog 接口，几毫秒重置。
- ArgaYC Spring 26
Stateful twins of SaaS services. We tested its Datadog twin.SaaS 服务的有状态替身。我们测过它的 Datadog 替身。
- Coarena · Ooak DataYC S26
Blind arenas for computer-use agents; evals inside real business workflows.电脑操作代理的盲测竞技场；在真实业务流程里做评测。
How we'd start我们怎么起步Twins of the 2–3 services agents touch most, plus scenarios and strict checks.先做代理最常碰的两三个服务的替身，加上场景和严格的检查。
Who pays谁付钱Agent teams, $1,250–3,500 a month (Arga’s tiers).做代理的团队，每月 1,250–3,500 美元（Arga 的价位）。
12RL environmentsRL 训练环境
Sales销售Engineering工程
What the companies do公司在做什么
- EdotEnvYC S26
Quant-research environments; markets keep raising the bar.量化研究环境；市场会不断抬高难度。
- Maingen · Cuebench · Standard MachinesYC S26
Industrial operations, scientific reasoning, chip design.工业运营、科学推理、芯片设计。
- Preference Modela16z
AI-research tasks, hardened against reward hacking.AI 研究类任务，专门防模型钻空子。
How we'd start我们怎么起步One domain, one environment with a strict grader. Show it separates models.一个领域、一个判卷严格的环境，证明它能把模型分出高下。
Who pays谁付钱About ten frontier labs: few buyers, big contracts, repeat buys as environments go stale.十来家大模型实验室：买家少、单子大，环境会过时，所以会复购。
EWhat YC asks for nextYC 下一批想要的
Fall 2026 Requests for Startups2026 秋季创业需求清单
13Multiplayer agents多人共用的代理
Product产品
What the companies do公司在做什么
- YC requestAaron Epstein
Teammates drop into the same live agent session to watch, steer or take over.同事能随时进到同一个正在跑的代理会话里，旁观、纠偏或接手。
- a16zBig Ideas 2026
“The collaboration layer becomes the moat.”“协作层会变成护城河。”
How we'd start我们怎么起步A shared live session for one team workflow, like support or code review.先给一种团队流程做共享会话，比如客服或代码评审。
Who pays谁付钱Teams, priced per seat.团队，按席位收费。
14Self-maintaining APIs自己维护的 API
Engineering工程Sales销售
What the companies do公司在做什么
How we'd start我们怎么起步Partner with one API provider about to ship a breaking change.找一家正要做破坏性升级的 API 公司合作。
Who pays谁付钱API companies, to cut the cost of migration support.API 公司，为了省掉迁移支持的成本。
15A cloud for small software小软件的云
Product产品Engineering工程
What the companies do公司在做什么
How we'd start我们怎么起步One-click deploy for agent-built tools inside a company, with sign-in and sharing.公司内部一键部署代理写的小工具，自带登录和分享。
Who pays谁付钱Non-engineers inside companies; teams pay per seat.公司里不写代码的人；按团队席位收费。
16An OS for the physical world物理世界的调度系统
Sales销售
What the companies do公司在做什么
- YC requestCharlie Warren
Route each job to an agent, a robot or a person: construction, maintenance, fleets.把每个活派给代理、机器人或人：工地、维修、车队。
How we'd start我们怎么起步One industry, like facility maintenance, starting with the work agents can already do.挑一个行业，比如设施维修，先从代理已经能做的那部分活开始。
Who pays谁付钱Field-service businesses. Long sales, hard to rip out.现场服务类企业。销售周期长，但用上就很难换。
17Consumer agents面向十亿人的代理产品
Product产品
What the companies do公司在做什么
- YC requestRaphael Schaad
Agents are good, and soon cheap, enough for consumer apps: getting around, learning, health, money, friendships.代理已经够好、很快也够便宜，可以做消费级应用：出行、学习、健康、理财、交友。
How we'd start我们怎么起步One everyday errand a person would hand to an assistant.挑一件普通人愿意交给助理去办的日常事。
Who pays谁付钱Consumers. Distribution decides everything.普通消费者。成败全看获客。
By weight按偏重看
Directions with two weights appear twice有两种偏重的方向会出现两次
Engineering工程01 Agent computers代理的电脑 · 02 Durable execution持久执行 · 03 Inference cost推理成本 · 11 Test sandboxes and evals测试沙盒与评测 · 12 RL environmentsRL 训练环境 · 14 Self-maintaining APIs自己维护的 API · 15 A cloud for small software小软件的云
Product产品04 Observability可观测 · 05 Identity身份 · 08 Memory记忆 · 09 Tools and data工具与数据 · 10 Software built for agents为代理而建的软件 · 11 Test sandboxes and evals测试沙盒与评测 · 13 Multiplayer agents多人共用的代理 · 15 A cloud for small software小软件的云 · 17 Consumer agents面向十亿人的代理产品
Sales销售05 Identity身份 · 06 Security and control安全与管控 · 07 Payments支付 · 10 Software built for agents为代理而建的软件 · 12 RL environmentsRL 训练环境 · 14 Self-maintaining APIs自己维护的 API · 16 An OS for the physical world物理世界的调度系统
02 · SimpleFunctions, reborn as an RL environment lab02 · 把 SimpleFunctions 改成 RL 环境实验室
Every market is an exam.
The world grades it.每个市场都是一道题，
世界来判卷。
SimpleFunctions, rebuilt as a lab that turns prediction-market data and trading into training environments for AI labs: what we would build, how each one grades, what we already have, who pays, and the first moves.把 SimpleFunctions 改成一家 RL 环境实验室，用预测市场的数据和交易给 AI 实验室做训练环境。下面是：做哪些环境、各自怎么判卷、已有的家底、谁付钱、怎么起步。
15,700+
Kalshi events settled last week, not counting parlays个 Kalshi 事件上周结算（不算串关）
2,100+
Polymarket events closed last week个 Polymarket 事件上周结束
1M+
settled markets with outcomes in our archive个已结算市场和结果，在我们的存档里
15 min
between world snapshots, recorded since April 2026一份世界快照，2026 年 4 月起
Week of Oct 13–20, 2026, counted from the two public APIs. We stopped paging early, so both numbers are floors. Most Kalshi events are crypto price ladders.按 2026 年 10 月 13–20 日两家的公开 API 数的。没有翻完，所以两个数都是下限。Kalshi 的事件里多数是加密货币的价格档位。
01
Grader built in自带判卷
The market is the question, its close date is the clock, settlement is the grade. No labelers.市场就是题目，到期日就是时钟，结算就是答案。不用人出题打分。
02
Restocks itself自己补货
Thousands of new questions settle every week. a16z: “Even the best environments have a shelf life.”每周都有上千道新题结算。a16z 说过：“再好的环境也有保质期。”
03
Genuinely hard真的难
Our April baseline couldn't significantly beat the market price at predicting settlement. For an environment, that's the point.SF 4 月做的基线模型，预测结算时没能显著赢过市场价。对训练环境来说，难正是卖点。
The environments要做的环境
Seven, in two groups; two to build first七个，分两组；先做两个
Environment · priority环境 · 优先级What the agent does代理做什么How it's graded怎么判卷Data it uses · skill it trains用到的数据 · 练的能力
A
Judgment判断
E1Forecast单题预测
Build first先做
What the agent does代理做什么At a sealed time t, get one question. Research with tools that only see up to t: prices, rules, the world snapshot, news signals. Then give a probability.在封存时间 t 拿到一道题，用只看得到 t 之前的工具查资料：价格、规则、世界快照、新闻信号。然后给出一个概率。
How it's graded怎么判卷Log score on the settlement, plus credit for beating the market price at t. Questions priced under 5¢ or over 95¢ at t are dropped.按结算算对数分，跑赢 t 时刻的市场价另加分。t 时刻价格低于 5¢ 或高于 95¢ 的题不出。
Data it uses · skill it trains用到的数据 · 练的能力Settlements, prices, world snapshots. Trains calibration and research against a deadline.结算、价格、世界快照。练校准，和限时调研。
What the agent does代理做什么Follow one market from open to close, waking at each new snapshot to revise the probability.从开盘跟到收盘，每来一份新快照就醒来，修正一次概率。
How it's graded怎么判卷Log score averaged over the whole path, so being right early earns more. Nothing after the current step is visible.对整条路径平均算对数分，越早判断对得分越多。当前这一步之后的东西一概看不到。
Data it uses · skill it trains用到的数据 · 练的能力The 15-minute snapshot history. Trains updating beliefs over days, the core problem of long-running agents.15 分钟一份的快照历史。练跨天更新判断，这正是长期运行的代理的核心难题。
What the agent does代理做什么A news signal lands at t. Say which markets will move in the next 24 hours, and which way.t 时刻来了一条新闻信号，说出接下来 24 小时哪些市场会动、往哪边动。
How it's graded怎么判卷Scored on the actual 24-hour moves, with over ten thousand samples a day. Our April LightGBM baseline: Brier 0.229, against 0.250 for a coin flip.按实际 24 小时的价格变动打分，每天上万个样本。SF 4 月的 LightGBM 基线：Brier 0.229，抛硬币是 0.250。
Data it uses · skill it trains用到的数据 · 练的能力News signals, world snapshots, price history. Trains cause-and-effect reasoning about the real world.新闻信号、世界快照、价格历史。练对现实世界的因果推理。
E4Resolve判结算
Build first先做
What the agent does代理做什么Read a market's resolution rules and the evidence after it closed, then rule how it settles, edge cases included.读一个市场的结算规则和收盘后的证据，判定它怎么结算，边界情况也算。
How it's graded怎么判卷Exact match with the official settlement: deterministic, no noise. The evidence never contains the settlement itself.和官方结算结果对答案：确定性判卷，没有噪声。给的证据里绝不包含结算结果本身。
Data it uses · skill it trains用到的数据 · 练的能力Rules text, settlements, post-close news. Trains careful reading of contract-style rules.规则原文、结算结果、收盘后的新闻。练读懂合同式的条款。
B
Trading交易
What the agent does代理做什么See 200 open markets at t. Pick the 10 most mispriced and say which side to take.t 时刻给 200 个还在交易的市场，挑出最错价的 10 个，说清买哪边。
How it's graded怎么判卷Realized edge of the picks: settlement minus entry price, after fees.按挑中的题的实际优势打分：结算价减买入价，扣掉手续费。
Data it uses · skill it trains用到的数据 · 练的能力Market-wide prices and settlements. Trains triage: deciding where attention pays.全市场某一时刻的价格和结算。练分诊：判断注意力花在哪最值。
E6Consistency关联与跨所
Later以后
What the agent does代理做什么Linked questions (“wins the primary” vs “wins the nomination”), or one event listed on both venues. Find prices that can't all be right and propose a hedge.关联题（比如“赢初选”和“赢提名”），或同一事件在两家都上了。找出不可能同时成立的价格，给出对冲方案。
How it's graded怎么判卷Logic checks score instantly (exclusive outcomes can't sum past 100%); the hedge is scored when prices converge or settle.逻辑检查当场就能判（互斥结果加起来不能超过 100%）；对冲方案在价格收敛或结算时打分。
Data it uses · skill it trains用到的数据 · 练的能力Cross-venue pairs, event groups, thesis trees. Trains keeping many beliefs consistent at once.跨所配对、事件分组、论点树。练让一大堆判断彼此一致。
What the agent does代理做什么Run a paper bankroll for a simulated month: size positions, fill against recorded order books, pay fees.管一个模拟账户跑一个月：定仓位，按录下的盘口成交，付手续费。
How it's graded怎么判卷Return per unit of risk, worst drawdown, beating the closing price. 15-minute books suit daily decisions, not high-frequency trading.看风险调整后的收益、最大回撤、有没有跑赢收盘价。15 分钟一份的盘口适合按天做决策，不适合高频。
Data it uses · skill it trains用到的数据 · 练的能力Order books, settlements, fees, and SF's order code. Trains taking risk under real constraints.盘口、结算、手续费，加上 SF 的下单代码。练在真实约束下承担风险。
What we'd sell卖什么
Product产品What's inside里面有什么Who buys谁买How it's priced怎么收费
P1
Environment packs环境包
What's inside里面有什么Tasks, graders and the sealed-time tool server, shipped as containers. The open versions follow Prime Intellect's verifiers format.题目、判卷器和封存时间的工具服务器，打成容器交付；开放版本按 Prime Intellect 的 verifiers 格式发布。
Who buys谁买Frontier-lab post-training teams大模型实验室的后训练团队
How it's priced怎么收费Annual license per environment. The price gets set in the first lab conversations.每个环境按年授权。价格在头几次跟实验室谈的时候定。
P2
Live feed实时题库
What's inside里面有什么Every week, fresh questions that settle after a given model's training cutoff, with graders.每周一批新题，都在某个模型训练截止之后才结算，附判卷器。
Who buys谁买Labs, for training and pre-release evals; eval groups实验室（训练和发布前评测）、评测机构
How it's priced怎么收费Monthly subscription. This is the recurring revenue.按月订阅，这是持续收入。
P3
Custom domains定制领域
What's inside里面有什么Environments built to a lab's spec: elections, macro, sports, weather.按实验室的需求定制：选举、宏观、体育、天气。
Who buys谁买Labs with a specific capability gap某项能力有缺口的实验室
How it's priced怎么收费Project fee按项目收费
P4
Public leaderboard公开排行榜
What's inside里面有什么Monthly scores for major models, on Polymarket questions only (Kalshi's terms don't allow it).每月给主流模型打分，只用 Polymarket 的题（Kalshi 的条款不允许）。
How it's priced怎么收费Free. It's how labs find us.免费，是让实验室找到我们的入口。
Who pays谁付钱
Customer客户What they use it for拿来干什么What they'd buy会买什么How we reach them怎么够到
Frontier labs前沿实验室
OpenAI, Anthropic, Google DeepMind, Meta, xAIOpenAI、Anthropic、Google DeepMind、Meta、xAI
What they use it for拿来干什么Forecasting and research skills in post-training; release evals后训练里的预测和调研能力；发布前评测
What they'd buy会买什么Environment packs and the live feed环境包和实时题库
How we reach them怎么够到Warm intros, the public leaderboard, a short paper熟人引荐、公开排行榜、一篇短论文
Open-model labs开源模型团队
Qwen, DeepSeek, MistralQwen、DeepSeek、Mistral
What they use it for拿来干什么The same skills on a smaller budget; open benchmarks matter more to them同样的能力，预算更小；更看重公开基准
What they'd buy会买什么The live feed and open environment versions实时题库和开放版环境
How we reach them怎么够到The leaderboard; Prime Intellect's Environments Hub排行榜；Prime Intellect 的 Environments Hub
Trading and research agent builders做交易和研究代理的公司
What they use it for拿来干什么Testing an agent before it touches real money代理碰真钱之前先测
What they'd buy会买什么Scan and Portfolio environments找错价和模拟账户环境
How we reach them怎么够到Prediction-market developer communities预测市场的开发者社区
Eval groups and academia评测机构和学术界
What they use it for拿来干什么Benchmarks and papers做基准、发论文
What they'd buy会买什么Little, but they bring citations买得少，但带来引用
How we reach them怎么够到Open data releases, workshops开放数据、研讨会
Kalshi and PolymarketKalshi 和 Polymarket
What they use it for拿来干什么Partners more than buyers: we need their data license更像合作方而不是买家：我们需要它们的数据授权
How we reach them怎么够到Directly, as part of the license request直接谈，和授权一起谈
What we already have我们已有的家底
From SimpleFunctions’ own records来自 SimpleFunctions 自己的记录
Asset家底Size规模Status状态Feeds喂给
Market recorder录制管线
Kalshi and PolymarketKalshi 和 Polymarket
Size规模About 53K markets tracked; prices and order books every 15 minutes跟踪约 5.3 万个市场；价格和盘口每 15 分钟一份
Status状态Off关着
Off since the September shutdown; the code is intact.9 月关停后没在跑，代码都在。
Settlements结算语料
Size规模1M+ settled markets with their outcomes100 多万个已结算的市场和结果
Status状态Have有
In the database, which was downsized, not deleted.在数据库里；数据库降了配置，没删。
World snapshots世界快照
Size规模One every 15 minutes: markets, indicators, signals; a daily export on Hugging Face每 15 分钟一份：市场、指标、信号；每天一份导出在 Hugging Face
Status状态Verify待核实
Archived daily from Apr 18. Expect Apr 18–Sep 28, plus Oct 6–12.4 月 18 日起每天归档，预计覆盖 4/18–9/28，外加 10/6–10/12。
News signals新闻信号
Size规模0.7M rows by mid-May: news our heartbeat picked up, every 15 minutes5 月中已有 70 万行：心跳任务每 15 分钟抓到的新闻
Status状态Verify待核实
In the database; date range to check.在数据库里，时间范围待查。
Order books盘口
Size规模Every 15 minutes; 4M rows by mid-May每 15 分钟一份；5 月中已有 400 万行
Status状态Verify待核实
In the database; date range to check.在数据库里，时间范围待查。
Cross-venue pairs, thesis trees跨所配对、论点树
Size规模Kalshi–Polymarket price gaps; causal trees linking theses to marketsKalshi 和 Polymarket 之间的价差；把论点连到市场的因果树
Status状态Have有
In the database.在数据库里。
Agent toolset代理工具集
Size规模76 tools under one contract for the CLI, SDK and agent76 个工具，命令行、SDK 和代理共用一份契约
Status状态Have有
Needs an “as of” time parameter to become the action space.加一个“截至某时”的参数，就是环境的动作空间。
Order execution下单代码
Size规模Kalshi and PolymarketKalshi 和 Polymarket
Status状态Have有
The code is intact.代码都在。
Feeds喂给E7 (simulated fills)E7（模拟成交）
Baselines and a benchmark基线和基准
Size规模April LightGBM on Hugging Face; World Awareness Bench: Claude Haiku 4.5 scored 2.3% alone, 70.5% with an 800-token world state4 月的 LightGBM 已发到 Hugging Face；World Awareness Bench：Claude Haiku 4.5 单独答对 2.3%，加上 800 token 的世界状态后 70.5%
Status状态Have有
Published.已公开。
Feeds喂给E3; proof we ship evalsE3；证明我们做过评测
Where the work is力气花在哪
Engineering工程Heavy重Sealing time so tools never see the future, graders that can’t be gamed, a recorder that never misses a day.把时间封死，工具看不到未来；判卷器钻不了空子；录制一天都不能断。
Sales销售Heavy重About ten buyers, and every deal is a relationship. One of the three of us has to own the lab conversations.买家就十来家，每一单都靠关系。我们三个人里得有一个专门负责跟实验室谈。
Product产品Light轻Packaging and docs: containers, task formats, one clear spec sheet per environment.打包和文档：容器、题目格式，每个环境一页清楚的说明。
How we would start怎么起步
Step步骤What we ship交付什么Done when过关标准Effort工作量
01
Recorder back on重开录制
What we ship交付什么Capture only: prices, order books, settlements, news signals, straight to R2只录、不做产品：价格、盘口、结算、新闻信号，直接存进 R2
Done when过关标准Seven straight days captured with no gaps连续 7 天录满，不断档
Effort工作量1–2 sessions1–2 个 session
02
Inventory盘点存货
What we ship交付什么A coverage table: every table by month, with the gaps marked一张覆盖表：每张表按月列出，标出断档
Done when过关标准We know which environments the archive can carry知道存档能撑起哪些环境
Effort工作量1 session1 个 session
03
Environment v0第一版环境
What we ship交付什么Forecast and Resolve: the sealed-time tool server, task generator, graders and filters单题预测和判结算：封存时间的工具服务器、出题器、判卷器、过滤规则
Done when过关标准Four models' scores spread apart, and a red-team agent can't game the grader四个模型的分数拉得开，红队代理也钻不了判卷器的空子
Effort工作量3–5 sessions3–5 个 session
04
Baselines and leaderboard基线和排行榜
What we ship交付什么Scores for Claude, GPT, Gemini, an open model and our LightGBM, on Polymarket questionsClaude、GPT、Gemini、一个开源模型和我们的 LightGBM 的分数，只用 Polymarket 的题
Done when过关标准The leaderboard is public排行榜公开
Effort工作量1–2 sessions1–2 个 session
05
Lab conversations找实验室聊
What we ship交付什么Results in front of 3–5 labs拿着结果去见 3–5 家实验室
Done when过关标准One paid pilot, or a clear no and the reason拿到一个付费试点，或者一个明确的拒绝和原因
Effort工作量People, not sessions靠人，不靠写代码
06
Data rights数据授权
What we ship交付什么A license conversation with Kalshi; a written check of Polymarket's terms和 Kalshi 谈授权；书面核实 Polymarket 的条款
Done when过关标准Written permission, or a Polymarket-only plan拿到书面许可，或者定下只用 Polymarket 的方案
Effort工作量Alongside 03–05和 03–05 同时进行
Who else is here还有谁在做
Who谁What they do在做什么Overlap跟我们重叠的地方How we differ我们的不同
EdotEnvEdotEnv
YC S26
What they do在做什么Quant-research environments for labs and academic groups给实验室和学术团队做量化研究环境
Overlap跟我们重叠的地方Markets as a testbed that keeps getting harder都把市场当成越来越难的考场
How we differ我们的不同Real-world events over time, not price data alone; and we have live-captured snapshots我们做随时间推进的现实事件，不只是价格数据；而且有当时实时录下的快照
Preference ModelPreference Model
a16z
What they do在做什么Environments for AI-research work, hardened against reward hackingAI 研究类工作的环境，专门防钻空子
Overlap跟我们重叠的地方The same buyers买家一样
How we differ我们的不同A different domain领域不同
Environment-first startups专做环境的创业公司
Mechanize, Fleet AI, HUD, Veris AI, Plato, Bespoke LabsMechanize、Fleet AI、HUD、Veris AI、Plato、Bespoke Labs
What they do在做什么Environments across many domains跨很多领域做环境
Overlap跟我们重叠的地方Any of them could add a forecasting domain谁都可能加一个预测领域
How we differ我们的不同Our sealed history can't be rebuilt after the fact我们封存的历史，事后补不出来
Data vendors adding environments加做环境的数据公司
Scale AI, Surge AI, Mercor, TuringScale AI、Surge AI、Mercor、Turing
What they do在做什么Human-labeled data, now environments too人工标注数据，现在也做环境
Overlap跟我们重叠的地方They have the lab relationships we don't他们有我们没有的实验室关系
How we differ我们的不同Our grader is the world, not labelers我们的判卷来自真实世界，不靠标注员
Academic prediction-market evals学术界的预测市场评测
Prediction Arena, PredictionMarketBenchPrediction Arena、PredictionMarketBench
What they do在做什么Benchmarks: models trade real money, or backtest on Kalshi data基准测试：模型拿真钱交易，或在 Kalshi 数据上回测
Overlap跟我们重叠的地方The same idea, used for evaluation同样的思路，用在评测
How we differ我们的不同We sell training environments and a fresh weekly feed我们卖训练环境和每周更新的题库
Risks风险
Risk风险What could go wrong会怎样What we do怎么办
Data rights数据权利
What could go wrong会怎样Kalshi's terms ban redistribution (our May 2026 legal review).Kalshi 的条款禁止转售（SF 今年 5 月的法务调研）。
What we do怎么办Ask for a license; use only Polymarket for anything public.去谈授权；凡是公开的东西只用 Polymarket。
Lab access进不了实验室
What could go wrong会怎样About ten buyers, and no warm intros yet.买家就十来家，现在还没人引荐。
What we do怎么办Leaderboard and a short paper first, then ask for intros.先有排行榜和短论文，再找人引荐。
Leakage答案泄露
What could go wrong会怎样A model may have seen the outcome in pretraining.模型可能在预训练里见过结果。
What we do怎么办Use only questions that settle after each model's training cutoff.只用在该模型训练截止之后才结算的题。
Noisy rewards奖励噪声
What could go wrong会怎样A 70% forecast is still wrong 30% of the time.70% 的预测也有 30% 会错。
What we do怎么办Proper scoring rules, many questions, the market price as the baseline.用恰当评分规则、大量题目，以市场价为基准。
Reward hacking被钻空子
What could go wrong会怎样Near-certain prices, leaked settlement text, peeks at future snapshots.几乎确定的价格、泄露的结算文字、偷看未来的快照。
What we do怎么办Filters, sealed tools, and a red-team agent before every release.过滤规则、封存的工具，每次发布前先让红队代理打一遍。
Short history历史太短
What could go wrong会怎样About five months of sealed snapshots.封存的快照大约只有五个月。
What we do怎么办Turn the recorder on now. Forecast and Resolve need little history.现在就重开录制。单题预测和判结算对历史长度要求不高。
Few buyers买家集中
What could go wrong会怎样Revenue rests on a handful of labs.收入压在少数几家实验室身上。
What we do怎么办The live feed makes revenue recurring; agent builders are the second market.实时题库让收入持续；做交易和研究代理的公司是第二市场。