当可靠的代码变成了偶尔发疯的OpenClaw，我们未来的工作范式变迁

Summary

This Keji Luandun episode uses hands-on Open Claw deployment to argue that AI agents change software from a predictable command executor into a probabilistic coworker with local files, accounts, browsers, memory, skills, and unsafe edge cases. The hosts connect Mac mini deployment, high remote-model cost, Kimi routing, scheduled-task drift, prompt injection, memory lock-in, MCP-style platform access, and AI coding into a broader shift from “by you” computer use toward “with you” agent management. Its main wiki contribution is Probabilistic Software: useful agent systems need bounded permissions, recoverable harnesses, deterministic subtools, verification, and human judgment because local autonomy amplifies both capability and failure.

Key Claims

• Open Claw is useful for low-risk, one-off, human-reviewed tasks such as checking flights or sorting information, but the episode is skeptical of letting it run unattended deterministic routines after seeing scheduled workflows and configuration files drift.
• Local Agent Execution is the attraction and the danger: the agent can use local browser state, logged-in accounts, files, iCloud-like data, and password-manager plugins in ways cloud agents usually cannot.
• Open source does not remove the core risk because the underlying model can still misunderstand, hallucinate, mutate its own environment, or follow injected instructions from a webpage or third-party skill.
• Docker or a virtualized environment is only a partial boundary if the user mounts important directories, tokens, accounts, or outbound network access into the agent’s workspace.
• The work pattern shifts from reliable software operated “by you” to uncertain collaboration “with you”: the user writes clearer tasks, maintains memory and AI Skills, gives feedback, reviews output, and manages agents like digital employees.
• Claude Code-style memory files and reusable skills show how advanced users may “train” agents through project norms, testing methods, product-design routines, and post-task experience updates.
• Enterprise agents will likely remain closer to fixed, auditable processes with selected skills than to fully autonomous personal agents because companies need reproducibility and control.
• MCP-like access could become an AI-era service interface: if platforms such as Meituan expose callable ordering or local-service actions, assistants such as Doubao or Yuanbao can route demand to them; if they do not, they may become less visible in AI workflows.
• AI service entry can reduce browsing and shift choice power to the assistant, as in the hosts’ milk-tea example where the user sees fewer options than in the traditional app interface.
• Good AI interaction should include clarification and triage, not only execution; the hosts contrast a medical-style assistant that asks follow-up questions with tools that mechanically complete a vague request.
• Persistent memory can become user lock-in because accumulated conversations, preferences, and task habits make switching agents costly; the hosts also speculate about a third-party “one memory” layer.
• AI coding makes one-off utilities easier, such as a Markdown-to-Word HTML tool, but debugging, answer checking, context loss, and repeated bug regressions keep AI Coding Verification and engineering structure central.
• The hosts warn of skill discontinuity: if future code is generated from scratch by agents, fewer people may develop the architecture, reading, and low-level practice that later expert review requires.
• Practical safety advice is to experiment on an isolated device, grant only narrow directories, avoid credit cards and high-risk accounts, and stop exposed services after use.

Key Quotes

“可靠的代码变成了偶尔发疯的OpenClaw” — the episode’s core metaphor for AI-driven software.

“by you” / “with you” — the hosts’ contrast between tool operation and agent collaboration.

“基模、应用、技能、数据” — the product stack the hosts use to describe future AI competition.

Connections

• Open Claw — central local-agent case.
• Probabilistic Software — concept added by this source for uncertain agent-mediated software behavior.
• Local Agent Execution and Agent Permission Boundaries — local data access, password-manager exposure, Docker limits, and least-permission guidance.
• Agent Harness, Persistent Agent Memory, and AI Skills — memory files, skills, task feedback, and recoverable execution loops.
• Human-Agent Collaboration and Human Judgment Under AI — “with you” agent management, clarification, and final responsibility.
• AI Coding Verification, Vibe Coding, and AI Engineering Thinking — temporary tool building, context loss, debugging, and skill-transfer concerns.
• Model Routing Cost Control, Model Workflow Fit, and Kimi — high remote-model cost and cheaper model routing in hands-on OpenClaw use.
• Model Context Protocol, Agent-Facing Interfaces, and AI Assistant Service Entry — platform visibility and service-fulfillment interfaces.
• Agentic Commerce, Meituan, Doubao, and Yuanbao — milk-tea ordering, AI recommendation power, and platform incentive questions.
• MiniMax — low-friction HTML deployment example where platform safety checks modified generated scripts.

Contradictions

• No direct contradiction with prior wiki content. The source reinforces existing Open Claw, Agent Permission Boundaries, Local Agent Execution, Persistent Agent Memory, AI Skills, and AI Coding Verification themes.
• It qualifies optimistic OpenClaw readings by stressing that local autonomy is still poorly suited to unattended, high-impact, or enterprise-deterministic work unless the surrounding harness, permissions, testing, and audit surfaces improve.
• It also adds tension to agentic-commerce enthusiasm: assistant-led ordering can lower friction, but it may also reduce user choice visibility and shift distribution power toward whichever assistant or platform controls recommendations.