“新建对话”的消失

Current hypothesis

个人 Agent 的成立意味着 session/conversation 会逐渐从用户面对的一级抽象退居系统内部。真正值得研究的是用户是否还需要承担“新建对话、在哪条对话继续、何时重开、哪些内容应该被记住”的管理责任。

早期 conversation 同时承载 state、context、task、time、UI 和 identity。随着长期 memory、projects、tasks、artifacts、background runs、persistent agents 等对象逐渐独立，这些职责开始拆分。

更稳定的对象模型可能包括 Agent、Goal/Responsibility、Project、Task、Artifact、Event、Conversation、Run/Session、Memory。Personal Agent 的前台趋向一个稳定关系，后台仍可有多模型、多 session、多 sub-agent、多工具。Multi-agent 很可能增长，但不应要求用户管理大量 bot/session。

Conversation 仍会存在，更像 interaction stream；session/run 则主要承担执行隔离、恢复、审计、计费等系统职责。界面重心由“操作 conversation”逐渐转向“委托 responsibility”。

Research branches

• persistent personal agent 与传统 chat/session 的对象模型差异；

• Grok Bot、Muse、Dots、ChatGPT/Codex Work 等产品如何暴露或隐藏 session；

• cross-session orchestration、voice control 与长期 agent identity 的关系；

• projects/tasks/artifacts/events 是否逐渐替代 chat history 成为主要导航结构；

• 多 Agent 增长是否会发生在后台，而前台仍保持单一稳定主体。

Maintenance rule

保持观点化主线，产品案例只用于验证或反驳抽象，不做产品堆砌。标题暂定《“新建对话”的消失》，后续继续发散后再决定是否定稿。

Workshop deliverable — 2026-10-14: Product thesis and evaluation plan

Refined thesis. The disappearance of “New Chat” can be measured by how often people must manually choose which conversation contains the right context. A persistent agent could associate new requests with an ongoing responsibility, retrieve relevant evidence, execute through separate runs, and present results in a stable interface. Conversations remain useful for intentional context separation and review.

Minimal product design. Expose a named responsibility only for ongoing work: objective, current state, evidence or artifacts, and next decision. Transient questions need no permanent object. The agent can manage internal runs, retries, and handoffs. Present decisions and finished outputs in a unified attention view. Avoid inventing separate persistent objects where a view over existing state suffices.

Evidence. OpenAI workspace agents provide recurring shared work as an explicit product concept. Codex app-server still uses thread/start and thread/resume to manage execution. These examples support the separation of user responsibility and internal execution, without proving a general shift in user habits.

Falsifiable test

Compare a conventional conversation-history interface with a persistent-agent prototype on 12 tasks: four returning to old work, four concurrent or interrupted workflows, and four tasks where the user deliberately chooses a fresh context. Hold model, tool budget, and materials constant. Measure manual navigation, correct retrieval of task state, task completion, recovery after interruption, ability to trace decisions to evidence, and perceived control.

Predictions. Persistent responsibilities should reduce manual session management for long-running work; execution sessions should remain important in developer and audit interfaces; and an explicit fresh-context action should remain valuable for user control. A study showing that users reliably prefer explicit conversation management would weaken the thesis.

Article structure for 《“新建对话”的消失》: (1) the burden of choosing which chat to continue; (2) why conversation historically combined context, work, and execution; (3) separation of persistent responsibilities and temporary runs; (4) the cases where deliberate boundaries remain essential; (5) how to measure whether the new abstraction actually helps.

Open evidence gap: longitudinal user studies, context-retrieval reliability, and empirical comparisons of attention-oriented versus conversation-oriented interfaces. This is a research proposal, not a verified universal trend.

Draft development — 2026-10-15

进一步收紧的主张：《“新建对话”的消失》研究的对象应是“由用户手动创建和管理上下文容器”这一交互责任。聊天文本本身仍会存在，执行 session/run 也会继续承担隔离、恢复、审计与计费；变化发生在一级产品抽象：系统逐渐接管“该续哪条、是否需要新上下文、应该带入哪些历史、何时把一次交互归入长期工作”这些决定。

这使三个层次可以明确分开：Agent / Responsibility 是长期主体与状态；Conversation 是人与系统的 interaction stream；Run / Session 是内部执行单元。Projects、memory、tasks、artifacts 与 event-triggered work 正在把原先堆在 conversation 里的 durable state 拆出去。OpenAI Projects 已允许不同 chat 共享同一项目上下文与历史；Dreaming memory 让未来 conversation 从共享长期上下文开始；Work 与 workspace agents 可在云端持续执行或由 schedule/trigger 启动。[1][2][3][4]

Grok Bot 是当前最直接的产品化验证：其设计明确把主对象从 chat history 改成 persistent Bot roster，并将 chat 定义为与 Bot 交互的界面；同时保留 Routines、artifacts、独立 runtime 等对象。这个案例用于验证抽象方向，不作为文章组织主线。[5]

文章的关键判断：未来仍可能保留“开始一个全新上下文”的显式动作，但它应成为少数情况下的 boundary control，而不是每次开始工作前都要选择的默认容器。衡量这一演化的核心指标，是用户为 context/session 管理付出的显式操作是否持续下降，以及系统在自动续接时是否保持可纠错、可追溯、可隔离。