Podcast Atlas

用 Agent 动力学，和 40 个 Agents 一起为「人 + AI」做产品｜对谈 Slock.ai 创始人 RC

Source note Episode guide Original audio Topics: Technology

Summary

This 42章经 episode interviews RC, founder of Slock.ai, about why CLI returned as an important interface in the agent era, what he learned from building Kimi CLI, and why he left Kimi to build a collaborative environment for people and many agents. RC argues that future software has to serve both humans and agents: command-line tools, AI Skills, memory, channels, threads, shared documents, task claiming, and model diversity become organizational infrastructure rather than peripheral product details. The strongest new contribution is Agent Dynamics: once a seven-person team works with roughly forty agents, product design becomes a question of identity, context, shared learning, task ownership, team culture, and human review.

Key Claims

  • CLI is newly valuable because large models are text-first systems; an Agent-Optimized CLI should have concise inputs, clear docs and examples, stable output, dense information, and obvious success/failure signals.
  • Kimi CLI is presented as a command-line coding/general agent, but RC says the deeper asset was a reusable local Agent Harness that could later support SDK, Web UI, or VS Code surfaces.
  • RC says a coding agent can be derived from a simple agent loop plus a Bash tool, then improved by observing what the model cannot do and adding tools and prompt structure.
  • Stronger coding models may widen security pressure because attackers can search for vulnerabilities in banks, kernels, browsers, compilers, and critical software faster than defenders can patch.
  • Skill plus CLI lowers the human software-use burden: the person should know what capability a tool gives the agent, while the agent reads instructions, installs tools, and uses them.
  • Slock.ai is designed as a collaborative environment for people and multiple agents, addressing hard-to-track local sessions, disconnected findings, and team knowledge trapped in individual human-agent interactions.
  • RC argues that Build and Code are becoming orthogonal: non-programmers can build with agents, while experienced programmers still have an advantage in serious software because they can inspect and constrain the agent.
  • Programming education may shift from bottom-up foundations first to top-down building first: users prompt an agent into producing a working artifact, then learn front end, back end, deployment, database, and code concepts as complexity forces them to.
  • Slock’s internal case is a small team with about seven people and forty agents, including engineering, head-of-engineering, design, growth, and strategy roles.
  • RC contrasts a single all-purpose agent with a multi-agent route: humans still often want micro-control because current subagents can drift, and unrelated tasks should not all share one context.
  • In an Agent Marketplace, adopting someone else’s agent may be closer to forking its memory, context, and collaboration history than downloading a static app.
  • Long-running work still needs shared channels: agents can communicate through chat, databases, code, GitHub issues, or new tools, but foreground and background work both need a place where state is visible.
  • Slock has to design UX simultaneously for humans and agents: people see channels and unread messages, while agents see linear events, summaries, IDs, and context windows.
  • Message-based multi-agent systems need Agent Task Claiming so multiple agents do not all rush to do the same job after one instruction.
  • RC says agents can forget identity, such as failing to recognize that they are the named Alice in a channel, making identity and context refresh product issues.
  • Agent Dynamics includes culture-like effects: prompts that ask agents to complement each other encourage cooperation, while race-style competition can produce falsehoods, empty claims, and denigration of other agents.
  • RC sees model diversity as part of Slock’s advantage: different models and tools can play different roles, so an application company should not assume one model provider will own the whole surface.
  • Slock’s target user expanded from the One-Person Company or independent builder toward one-to-one-hundred-person teams that manage and interact with many agents.

Key Quotes

“Agent 动力学” — RC’s name for the behavior that appears when many agents interact in one work environment.

“7 个人和 40 个 Agent” — the source’s concrete team-scale example.

“Build 与 Code 已经变成正交关系” — RC’s claim that building with agents is separating from traditional coding skill.

Connections

Contradictions

  • No direct contradiction with prior wiki content. The source reinforces existing CLI, skills, harness, memory, and AI-first organization themes while adding a more explicit multi-agent sociology layer. The main tension is scope: older one-person-company material often frames agents as leverage for a solo operator, while this source says the value becomes clearer when one or more humans jointly manage a population of agents.