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 Episode summary Updated 2026-07-23 Tags: Podcast, Ai, Agents, Startup, Collaboration

Summary

This 42章经 episode interviews RC, founder of [[SlockAI|Slock.ai]], about why CLI returned as an important interface in the agent era, what he learned from building [[KimiCLI|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.
  • [[KimiCLI|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.
  • [[SlockAI|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.