Shared Company Agent Design
What it is
A shared company agent serves multiple people through common work surfaces and infrastructure. Its design must make conversational context, connected identities, and permission boundaries explicit, then introduce the system through owned, measurable workflows.
Why it matters
Sharing an agent can spread useful workflows across a team, but it also changes whose conversations the agent handles, which credentials it can use, and who may authorize an action. These are application and governance decisions even when the model and runtime are managed services.
Key points
- Treat conversation context as shared state. In Every’s account, Slack is the durable record from which a session can be reconstructed. The agent uses message routing and thread isolation to decide when to respond and avoid mixing separate conversations.
- Separate workflow identity from an individual’s login. Every describes scheduled work pointing to credential slots that can be replaced when access changes. It also assigns owners to skills and scheduled jobs, with an administrator path when an owner leaves. Credentials are encrypted server-side and kept out of prompts and conversations.
- Enforce permissions in the application. Every describes permission rules below the model and checking claimed consent against what the person actually wrote in Slack. Its experience also shows a cost of restrictive defaults: a migration made 24 skills private, and the team later restored access to nine. The account is a case study, not a universal authorization policy.
- Roll out one bounded workflow at a time. A small-team proposal recommends starting with a business unit or workflow, preparing the needed information, assigning ownership, and measuring an outcome. It suggests a small group of architects and practitioners supporting broader use through documented SOPs and skills.
- Centralize controls as the workflow earns them. The small-team proposal lists shared data, MCP services, API-key controls, skill versions, logs, and permissions as possible infrastructure. It favors workflow plus LLM as an initial step before long-running managed agents.
- Make trade-offs reviewable. The proposal explicitly accepts some compliance cost for speed in its small-team context. That is an assumption to assess against the actual data, regulation, and security needs, not a blanket recommendation.
Evidence across sources
| Source | Key claim | Relevance |
|---|---|---|
| Every — Why We Replaced Personal Agents With One for the Whole Company | The Every team describes redesigning conversation, connection, and permission assumptions for a shared Slack agent, including thread isolation, credential slots, code-level rules, and workflow ownership. | Detailed operational case for shared-agent behavior and governance. |
| AI Briefing 2026-10-22 Morning | A small-team adoption proposal emphasizes bounded workflows, prepared data, ownership, shared controls, staged adoption, and human-reviewed feedback; it also states its own compliance-speed trade-off. | Adds a rollout lens and a caveat about applying the proposal outside its stated context. |
Open questions
- When should a shared agent speak without being directly invited, and how should that boundary change as a team learns its behavior?
- How should credential slots and application-level permissions work across multiple workspaces or organizations?
- Which compliance and security requirements should block a proposed workflow before rollout?
Prompts for witness
- Which of your recurring workflows currently depends on one person’s credentials or private setup?
- If an agent acted in a shared channel, how would it establish whose request it is following and what that person may authorize?
- What is the smallest measurable workflow that could test shared value without broadening access prematurely?
Related
- Shared Brain First
- Zero Trust AI Agents
- Managed Agents Architecture
- Compound Engineering
- Single Master Agent Pattern
- Agent Login Agency — 凭证槽位与工作流身份分离的登录代理权设计
- Every and Notion AI Agents
Pending Jean review
- harness-engineering/shared-company-agent-design — ai_draft; suggested action: keep as a draft and judge the proposed ownership, credential, permission, and compliance model against one real team workflow before treating it as a general pattern.
2026-10-23 — Managed agents expose retention and approval trade-offs
来源:AI Briefing 2026-10-23 Afternoon;转述自 @Every。
Every 表示,团队用一周搭建 Claude Managed Agents 共享版本,把设计重点放在行为边界和连接器权限;凭证不向 Agent 暴露,共享 skills 的修改需人工批准。该部署也带来明确的平台代价:依赖 Claude,且对话历史持续保存,因此无法满足 zero-data-retention。它为已有的共享上下文、凭证隔离和应用层权限案例补充了托管平台的保留策略与迁移成本。
Counterpoints & Gaps
- 这是 Every 对自家部署的说明,不能据此推断其他托管 Agent 平台的权限或保留策略。
- 一周上线、行为设计与连接器控制的描述来自单一团队经验;迁移成本和技能审批效果尚无独立评估。
Sources (addendum)
Pending Jean review
- harness-engineering/shared-company-agent-design —
ai_draft; compare provider portability, conversation retention, and human approval for shared-skill changes against one real workflow before treating the managed-agent trade-off as general.
2026-10:Every 从私人 agent 到公司 agent 的演进
来源:AI Builders Digest 2026-10-20
Every 的完整播客解释了他们为何从「人人一个 OpenClaw 私人 agent」转向「一个公司级 agent」:私人 agent 对 tinkerer 有吸引力,但对运营者是安全与维护负担;共享 agent 让每次使用都变成组织资产,避免"记住哪个 bot 擅长什么"的开销;预测未来分裂为工作公司 agent + 家庭个人 agent 两个。
2026-10:Every 共享 agent 的组织扩散机制
Every 近一年实验的补充细节:组织 AI 采用的瓶颈是可见性而非能力——公共线程同时承担执行、演示、复盘、培训,但同一机制也决定泄露半径(数据分级边界要随 prompting in public 显式设计)。纠正进入可复用 skill 才是复利,skill promotion 需要评审条件:重复出现、跨输入有效、success check、owner、可撤销。
2026-10-22:Every 删除 Plus One 的完整复盘
第一人称复盘的增量细节(线程隔离、凭证槽位、权限规则进应用层与 24 个 skill 私有化事故本页 Key points 已记录,不再重复):
- 删除了 86,730 行个人 Agent 代码。个人 Agent 的三个隐含假设——单人对话、个人凭证、全权默认——在共享场景全部失效,这是重构的第一性理由。
- 权限规则写进代码而非提示词:模型无法被"说服"通融,代价是规则会"可靠地出错"(上述 24 个 skill 事故即此类);兜底准则是按"撤销难度"决定授权宽松度——越难撤销的操作默认越紧。
- CMA 经济学:$0.08/会话小时;修复长会话问题后总成本下降 39%;把迁离成本写进文档,让供应商锁定暴露为可讨论项。
- 行为回归案例:模型切换到 Opus 5 后回复长度从 280 字符涨到 880 字符——共享 Agent 对模型切换敏感,模型升级本身应按一次部署事件处理并纳入 eval。
2026-10-27 — Replit 与 Manus 把组织工作面向 Agent
来源:AI Briefing 2026-10-27 Afternoon;转述自 @amasad 与 Hidecloud。
Replit 的“Agent Habitat”主张把 Agent 接入组织知识和系统,同时明确访问范围、人工审批和错误影响边界;人仍负责目标、判断与问责。Manus 的 Mono 提供跨仓库统一入口、Agent 可读说明和受权限控制的线上日志,并用测试账号端到端运行和录屏验证结果。两者共同指向:组织要让 Agent 看见完成工作所需的上下文,也要让授权、审批与结果验证成为工作面的组成部分。共享 Session 可以承载协作,但仍需控制谁能查看或发言。
Counterpoints & Gaps
- Replit 的产出数据和未来预测来自公司自身文章;Manus 的流程与规模来自一线案例,均不能单独证明普遍效果。
- 不同组织的权限模型、数据分类和合规要求不同;Agent 可见的内容应按具体工作流评估。
Sources (addendum)
- AI Briefing 2026-10-27 Afternoon
- Amjad Masad — The Rise of the Self-Driving Company
- Hidecloud — 用 Manus 改造 Manus
2026-10-30:YC 的"梦境循环"组织自我改进实践
来源:增长黑客AI周报 EP72(Pete Koomen & Tom,YC)
YC 普通合伙人 Pete Koomen 在播客中介绍了 YC 的 AI 原生组织实践。通用 agent 每晚读取员工与 agent 的对话和会议记录,寻找改进点,写回内部 Postgres,形成"梦境循环"。"两句话描述技能"是典型案例:合伙人 Tom 写了初始技能,小组 office hours 的会议记录被 agent 学进去后,技能明显变好,Pete 认为它已超过自己。前提是全员完全可见、默认信任、共享组织大脑。成本量级为每年 $10k-$100k token。效果是组织内部可见杰文斯悖论——问题变得更繁多、更复杂。
Counterpoints & Gaps
- YC 是全员高密度的小团队,老板就在隔壁;一个 200 人公司必须先审视自己的信任基础是否支撑这种全员可见度。
- 梦境循环的效果描述来自 Pete Koomen 自述,没有独立的技能质量评估或对比数据。
Sources (addendum)
- 增长黑客AI周报 EP72 — YC 梦境循环与组织自我改进
Pending Jean review
- harness-engineering/shared-company-agent-design —
ai_draft; compare the proposed Agent Habitat and Manus Mono boundaries with one real workflow before treating them as a general organization pattern.