OpenClaw 之后，谁将定义主动式 AI 的新战场？｜对谈 AirJelly 黄柏特

Summary

This Shizilukou Crossing episode interviews AirJelly founder Huang Bote and angel investor Yihao about proactive, context-aware personal agents after the Open Claw wave. AirJelly’s thesis is that useful Proactive Agents need Intent Context, OS-Level Context, Persistent Agent Memory, and execution ability rather than only chat prompts, scheduled reminders, or generic push notifications. The discussion also connects agent products to startup strategy, privacy, AI Organization Design, and the 2026 investment window around agent apps, agent infrastructure, and hardware that captures context.

Key Claims

• AirJelly is positioned as an active context-aware partner that remembers what users do across tools and appears at the right moment to help.
• Huang Bote distinguishes AirJelly from Mycontext: Mycontext relied on periodic screenshots, while AirJelly tries to detect events, tasks, intent, and memory worth preserving.
• AirJelly uses the Enter key as an Intent Context trigger because IM, chatbot, and search workflows often express user intent at the moment of submission.
• The episode argues that broad proactive AI includes reminders, scheduled tasks, and products such as ChatGPT Pulse, but the stronger form combines explicit current intent with enough surrounding context to advance the user’s next step.
• AirJelly’s memory model organizes context into events and entities, then uses merge, time decay, vector search, and reranking to keep memory useful rather than treating full recording as equal-value history.
• Open Claw and Claude Code initially pushed the team toward task-engineering and simplified execution tooling, but those directions looked too easy for larger products to cover; AirJelly returned to context capture, memory, and recall as the harder product layer.
• Huang Bote treats memory as the core moat for 2C agents: once a user has built one to three months of personal context, migration becomes harder.
• AirJelly’s proactive behavior is intended to extend the current task rather than create extra information load; timing is adjusted through work state, app switching, dismissal, and other feedback.
• Privacy is framed as a necessary product constraint: the team says pictures and virtual context can stay local, communication should use end-to-end encryption, and PII should be desensitized.
• Yihao says QuickStone had already invested in proactive AI teams and sees 2026 opportunities in vertical but aggressive agent applications, agent infrastructure, and hardware that gives agents more context.
• The AirJelly team experiments with an AI-native organization pattern where agents for different teammates can coordinate in a group, but the founder rejects turning the product into employee monitoring software.

Key Quotes

“Cursor 重新定义了 Tab，而 AirJelly 想重新定义 Enter” — Huang Bote’s product metaphor for intent capture.

“不是发生的每件事都会成为历史” — the source’s distinction between full recording and meaningful memory.

“如果一个产品能靠 vibe coding 做到 60 分或 80 分，就可能不值得创业” — Huang Bote’s startup-bar argument.

“顺着任务延长线推动” — AirJelly’s desired proactive behavior, contrasted with distracting information expansion.

Connections

• AirJelly — central proactive personal-agent product.
• Huang Bote — founder explaining product definition, memory architecture, open-source background, privacy posture, and startup timing.
• Yihao and QuickStone — investor and firm framing proactive AI, agent infrastructure, and context hardware as investment directions.
• Mycontext — predecessor project that captured screenshots and shaped AirJelly’s context thesis.
• Open Claw, Claude Code, Manus, ChatGPT, and Cursor — reference products used to define what current agents and chatbots still miss.
• Proactive Agents, Intent Context, OS-Level Context, and Persistent Agent Memory — core product concepts strengthened by the episode.
• Context Engineering, Agent Native Software, Agent Harness, and Agent Permission Boundaries — broader agent-design vocabulary connected to context capture, execution, skills, and privacy.
• Human-Agent Collaboration — AirJelly’s version moves away from one-shot prompting toward an agent that sees task state and intervenes when useful.
• AI Organization Design and AI Workforce Monitoring — organizational theme raised by the team’s no-meeting experiments and anti-monitoring boundary.
• WeChat and ByteDance — context environments and founder background relevant to the product path.

Contradictions

• No direct contradiction with prior wiki content. The source reinforces the existing Paperboy and Open Claw themes that durable agent value depends on context, memory, timing, permissions, and product form rather than model calls alone.
• The source adds a useful tension inside Proactive Agents: proactive behavior is not inherently good, because generic prompts and curiosity expansion can increase cognitive load unless grounded in current intent and task state.