Podcast Atlas

探秘 Claude Code，搞懂 Agent Harness｜对谈来新璐

source Updated 2026-07-06 Tags: Podcast, Ai, Agents, Infrastructure, Coding

Summary

This Shizilukou Crossing episode interviews Lai Xinlu of Share AI about Agent Harness design, using Claude Code and Learn Claude Code as the main teaching sample. Lai defines harness as the model-external system that gives an agent execution ability, context, state, memory, permissions, and orchestration. The discussion argues for model-aligned harness design: stronger models should get more context and action capacity with less brittle programmatic control, and CLI/Unix-like interfaces may be more agent-native than prompt-node or flow-graph frameworks.

Key Claims

  • Agent Harness can be understood as “everything outside the model”: tools, runtime environment, context, state, memory, permissions, and multi-agent governance.
  • Lai breaks harness design into three layers: execution ability, context/environment, and governance/orchestration.
  • The execution layer includes files, search, browser operation, language interpreters, CLI tools, code-registered tools, and MCP-like extensions, but the episode argues that CLI often outperforms newer abstractions because models have more pretraining exposure to Unix and shell patterns.
  • The context/environment layer includes system prompts, AI Skills, Persistent Agent Memory, context-window management, compression, and cross-agent handoff documents.
  • The governance/orchestration layer controls multi-agent roles, information access, and permissions, such as giving exploration agents read-only capabilities so they cannot mutate code or tests.
  • Claude Code is presented as an especially instructive harness because of its memory update hooks, AutoDream-like consolidation, markdown memory files, context compression, and task handoff mechanisms.
  • Good harnesses should fit the model’s operating logic and remain orthogonal to model improvement; a harness that over-controls the model may become a bottleneck as models improve.
  • K Computer represents Share AI’s lightweight Unix-style approach: a virtual computer built as data structures rather than a traditional Docker-like sandbox.
  • The entrepreneurial direction extends beyond coding tools into hybrid agent networking, agent payment, personalized model training, and agent-native organizations such as “zero-person companies.”

Key Quotes

“模型以外都是 Harness” — Lai’s compact definition of agent harness.

“模型才是 Agent” — the reason the episode emphasizes model capability before harness cleverness.

“更多 context、更少 control、更多 action 能力” — the preferred direction for Claude Code-style harness design.

“CLI is all you need” — the Unix/CLI-oriented interface thesis.

“最好的管理就是不要做管理” — the warning against context management that fights the model’s own operating pattern.

Connections

Contradictions