Podcast Atlas
source Episode summary Updated 2026-07-23 Tags: Podcast, Ai, Agents, Ai-First, Organization-Design, Software-Engineering

E238｜聊聊Harness时代AI-First的组织架构：从信任人到信任AI

Summary

This 硅谷101 episode interviews [[PeterCreo|Peter]], [[ChenKaiCreo|陈凯]], and [[ClarkCreo|Clark]] of Creo about Harness Engineering and [[AIFirstOrganization|AI-first organization design]]. The discussion argues that the next step after prompt and Context Engineering is not only better instructions, but an Agent Harness that lets agents run through real workflows, collect feedback, repair failures, and improve. Its strongest synthesis is organizational: Creo frames AI-first work as moving from people using AI tools toward AI driving production while humans provide architecture, value definition, judgment, and review.

Key Claims

  • [[PeterCreo|Peter]] defines Harness Engineering as a dynamic system problem around tooling, sandbox architecture, host-service interaction, security, startup time, latency, feedback, self-healing, and self-improvement.
  • The episode distinguishes [[AIFirstOrganization|AI-first organization design]] from ordinary AI tool adoption: [[ChenKaiCreo|陈凯]] says workflows and organization form have to be rebuilt around AI capability rather than adding AI on top of old handoffs.
  • Creo says its roughly 25-person company has 99% of code written by AI and can move from feature idea to A/B testing and rewrite within a day; this is a source-scoped internal estimate, not a general industry benchmark.
  • Faster implementation shifts the bottleneck from engineering delivery toward product choice, testing, quality control, market readiness, and go-to-market narrative.
  • [[PeterCreo|Peter]] describes agent-driven CICD, agent-driven bug triage, and low-risk autofixing PRs where different agents identify, assign, and repair frontend, backend, or agent-system issues.
  • The episode treats AI Coding Verification as a harness function: Playwright checks, AI-driven integration tests, rollout/fallback decisions, and human proof sit inside the production loop rather than after code generation.
  • [[ClarkCreo|Clark]] says go-to-market work is harder to evaluate than engineering because it faces humans and agents, and because market readiness remains a judgment call even when agents generate many candidate outputs.
  • The source argues that products may increasingly be consumed by agents, not only by humans, so SaaS value can move toward APIs, MCP-like access, permissions, and data structures that agents can inspect and prioritize.
  • [[ChenKaiCreo|陈凯]] frames organization redesign as a trust shift: companies have to decide when they can trust AI to plan, execute, and make recommendations, then build guardrails that make that trust auditable.
  • The guests argue that product-manager, engineer, designer, and marketer boundaries blur when AI absorbs implementation and coordination; generalists with architecture, product sense, implementation ability, and market judgment become more valuable.
  • The episode keeps humans in the loop at a higher level: architecture, security and latency judgment, value definition, demand direction, final review, and ethical boundaries remain human responsibilities.

Key Quotes

“99% 的代码由 AI 写” - Peter’s description of Creo’s internal engineering workflow.

“AI drive 公司方向和日常工作方式” - Chen Kai’s AI-first organization definition.

“2026 年自己没有写过一行代码” - Peter’s personal coding-role claim.

Connections

Contradictions

  • No direct contradiction found.
  • The source qualifies Agent Harness by emphasizing dynamic self-improvement and organization feedback, not only a static wrapper around tools and memory.
  • The source qualifies AI Coding Verification by showing a high-AI-code workflow that still depends on architecture, tests, rollout metrics, human proof, and permission boundaries.
  • The efficiency claims are Creo self-reports and should not be generalized to regulated enterprises, large legacy organizations, or nontechnical teams without additional sources.