OpenClaw 之后，我只想未来 3-6 个月的事情｜对谈 Sheet0 创始人王文锋

Summary

This 42章经 episode interviews Sheet0 founder 王文锋 / Wang Wenfeng about the post-Open Claw agent wave, especially the idea that [[CodingAgentAsUniversalActionLayer|coding agents can become the general action substrate]] for many non-programming workflows. Wang argues that AI Skills, file-system state, [[AgentOptimizedCLI|CLI]] surfaces, permissions, and feedback loops can let agents absorb work once packaged as [[AgentHarness|harnesses]], making the boundary between vertical agents, SaaS, and software development less stable. The second half uses Sheet0’s internal workflow to explain AI Managing AI: AI reads tasks from project management, develops and tests changes, opens GitHub PRs, and leaves humans mainly with product definition, taste, and final review.

Key Claims

• Open Claw and the earlier Minus agent wave are presented as continuity rather than a clean break: both proved new agent product forms, while OpenClaw made coding ability more visible.
• Wang treats AI coding as the model’s “dexterous hand”; in his stronger version, many agents become coding agents because code, files, CLIs, and tools let them act across domains.
• AI Skills can carry domain know-how when the work can be explained as procedures, acceptance criteria, tools, and examples; this is why Wang questions whether narrow vertical agents and traditional SaaS keep their old shape.
• SaaS is described as a historical way to scale expert workflows through UI, forms, dashboards, and records; Agentic Workflow shifts value toward agents that understand goals and instantiate expert work more directly.
• Long-horizon work is framed less as manual context packing and more as giving agents a computer-like environment where state, memory, progress, and errors are visible in files.
• Proactive Agents have a weak form, such as scheduled summaries or reminders, and a stronger form, where an agent understands business context, explores, reflects, and proposes action like a colleague.
• The source’s main product thesis is AI Managing AI: a meta agent can break down requirements, configure or call other agents, and close the loop through tests, screenshots, PRs, and human review.
• For Sheet0, the bottleneck moved from production speed to product definition: Wang says work that might once take months can now take weeks, so the scarce task is deciding what to build and what quality bar to hold.
• Agent Harness is described as the scaffold that lets strong agents operate in organizations: project context, team process, permissions, databases or read-only state, CLI access, and review feedback matter as much as the model call.
• Token spending becomes a management variable: Sheet0 spent about $20,000 on AI coding in the prior month, and Wang expects future customer segmentation to track token consumption and labor substitution rather than only headcount.
• Wang’s startup-method update is to reduce long-term prophecy: after overbuilding for five-to-ten-year problems, he now wants to follow real bottlenecks visible in the next three to six months.

Key Quotes

“AI coding 是大模型的灵巧手” — Wang’s explanation for why coding agents can become a general action layer.

“less structure, more intelligence” — shorthand for letting agents maintain context and workflow state themselves.

“预判为辅，跟随为主” — Wang’s revised startup method after over-weighting long-term predictions.

“只考虑三到六个月” — the title-level discipline for operating amid fast AI change.

Connections

• Sheet0 — startup whose product direction is shifting toward AI-managed engineering work.
• 王文锋 / Wang Wenfeng — founder explaining the OpenClaw interpretation, Sheet0 pivot, and short-horizon operating method.
• Open Claw, Claude Code, Codex, and Cursor — reference tools used to show coding-agent capability, adoption, and workflow leverage.
• Coding Agent As Universal Action Layer, AI Skills, Agent-Facing Interfaces, and Agent-Optimized CLI — mechanism by which coding agents extend beyond narrow programming tasks.
• Agent Harness, Agentic Workflow, Context Engineering, Persistent Agent Memory, and Agent Permission Boundaries — infrastructure and operating concepts needed for reliable agent work.
• AI Managing AI, Proactive Agents, Routine Agent Automation, and Long-Horizon AI — related agent-product patterns extended by the source.
• AI Native SaaS Threat, SaaS Trust Moat, Digital Employees, and Outcome-Based AI Pricing — SaaS and business-model consequences of agent-delivered work.
• AI Inference Cost Structure, Token Maxxing, Model Routing Cost Control, and One-Person Company — economics and organization implications of heavy agent usage.
• Human Judgment Under AI, AI Organization Design, and AI-First Organization — human role shift toward product taste, context, review, and final accountability.

Contradictions

• No direct contradiction with prior wiki content. The source strengthens existing Open Claw, Agent Harness, and Agentic Workflow themes, while adding a stronger skeptical view of vertical agents and SaaS than some enterprise-software pages; that tension is recorded as a scope difference rather than a settled conflict.