Podcast Atlas
concept Updated 2026-08-24

Agent Native Software

E249｜Token经济转点：OpenClaw、Hermes到本地自研的Agent进化之路 adds 东旭 / Dongxu’s direct test for agent-native businesses: if removing AI or the agent leaves the business basically intact, the product is not truly AI native or agent native. The episode applies this test to startup opportunities, arguing that infrastructure, memory, search, sandbox, collaboration harnesses, and agent-cloud services are more plausible near-term opportunities than thin applications that merely add an AI layer to old workflows.

Nikesh Arora: Mythos is Real, Analytical SaaS is Dead, and Google can be a $10T company adds Nikesh Arora’s enterprise-systems version. He argues that systems of record and systems of work will be reinvented as agents perform work behind or instead of human-facing UI, turning products such as Salesforce and Oracle into substrates for agent execution and Agent-Managed Audit Trails.

Agent-native software is software whose core substrate is an agent rather than a traditional app with an AI feature bolted on. In Vol. 161 从开发自己的 OpenClaw 聊起, Justin Yan and 自立 frame Open Claw this way: the surrounding code, tools, channels, UI, and AI Skills act as the agent’s hands, senses, and work environment, while the agent itself is what makes the product coherent.

Vol. 164 从苹果聊到软件未来：Agentic Software 真的要来了？ sharpens the adjacent Agentic Software definition. The hosts argue that simply adding an AI assistant, MCP-style tool access, or skills to existing software is not enough; the deeper change is decomposing products into callable abilities, dynamic interfaces, and human-agent review loops.

Vol. 165 做客声东击西：「龙虾」和 vibe coding 正如何改变我们的思维 adds a non-engineer perception shift. 徐涛 starts from an ordinary chat surface and only later realizes that the useful part of “小龙虾” is a programmable, layered system doing work behind the conversation. That makes agent-native software legible as a new way to package back-end routines, memory, and scheduled action for people who would not normally describe themselves as software builders.

OpenClaw 之后，谁将定义主动式 AI 的新战场？｜对谈 AirJelly 黄柏特 adds AirJelly as a context-first agent-native case. Huang Bote argues that task execution and orchestration alone can be copied quickly, while the harder product layer is giving the agent senses, memory, timing, and privacy-aware context. This reframes agent-native software as not only “what can the agent do,” but “what can the agent know safely and at the right moment.”

20 个问题，搞懂 OpenClaw：爆红机制、本质变化、创业机会 adds a packaging-centered view. 鸭哥 and 豪大 argue that Open Claw became legible because it combined IM Agent Interfaces, Local Agent Execution, memory, skills, and tool feedback into a product users could treat as a trainable assistant rather than a normal app with an AI button.

Vol. 170 Fable 5 重出江湖，GPT 仍需努力 extends the form into Token-Driven Software. Instead of only treating the agent as a worker behind fixed app screens, the hosts imagine software whose interface, interaction flow, and world behavior are generated at run time from context. This makes Model Routing Cost Control and AI Inference Cost Structure part of product design, because dynamic behavior can become expensive if every generated surface uses the top model.

Vol. 167 Token 如流水，Agent 似朝阳 adds the product-prototyping version. Open Claw and Hermes Agent-style agents can be configured in IM threads for article triage, translation, todo aggregation, and calendar/reminder synthesis, letting a builder test a product idea through conversation before deciding whether to engineer it as a stable app or skill.

141. Freda的投资札记第2集：Tokenmaxxing、把电机塞进蒸汽机、接力赛变篮球赛、孤独、人的连接 adds the enterprise-system redesign version. Freda / Friday argues that many AI CRM or ERP products still resemble old software with automation added. The stronger agent-native opportunity is to record previously invisible decision context, make systems persistent and real-time enough for agents, and redesign communication, permissions, memory, and workflow around nonhuman operators.

174: AI冲击企业软件巨头？与SAP原欣聊大模型to B的颠覆与边界 adds a conservative incumbent version through SAP. The source says SAP wants to make the application layer thinner and move from record system toward executable system through SAP Joule / Joule Work and agents, but it keeps ERP process, data, permissions, and localization as the governed substrate rather than treating agents as a full replacement.

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 adds Slock.ai as a multi-agent workspace form. The product is agent-native not because it adds an assistant to chat, but because channels, threads, documents, memory, task claiming, and review surfaces are designed around many agents as active participants in the workspace.

Key Claims

  • Agent-native software differs from AI-assisted software because removing the agent would remove the product’s reason to exist.
  • Agentic Software can include agent-native products, but it also describes how existing software may be rebuilt around Atomic Capability Services and agent-facing access.
  • The design center shifts from screens and static feature menus toward Agent Harness choices: tools, permissions, channels, triggers, memory, and feedback loops.
  • AI Skills become product surface because they define reusable capabilities and can sometimes be created or refined by the agent itself.
  • On-Demand Apps are one possible downstream form: the agent assembles or generates capabilities at the moment of need instead of exposing only prebuilt app functions.
  • Agent-native software increases the importance of Agent Permission Boundaries because broader action capacity also broadens failure and leakage risk.
  • Context capture and memory can be as much a product surface as tools and skills, especially for personal agents that need to act before being prompted.
  • Accessibility and entry point can be as important as raw capability: an IM surface plus local runtime can expose existing CLI-agent power to many more users.
  • Token-driven interaction broadens agent-native software from task execution into dynamic experience generation, but increases cost, latency, and quality-control requirements.
  • Agent-native prototypes can start as configured conversations, but durable products still need stable memory, permission boundaries, and deterministic pieces when repeated work becomes clear.
  • For non-technical users, agent-native software can reveal the programmatic structure behind work: chat becomes the surface for routines, state, memory, and tool execution.
  • Agent-native enterprise software may need to capture why decisions were made, who objected, what constraints mattered, and which approvals shaped the outcome, not only the final structured record.
  • SAP adds that incumbent enterprise software can move toward agent-native execution without abandoning the ERP substrate, especially when auditability and global compliance remain central.
  • Persistent, real-time systems can matter because agents may need to stay online, react to events, and maintain state rather than operate as one-shot assistants.
  • Agent-native collaboration software may need to be readable by two kinds of workers at once: humans scanning progress and agents consuming structured event history.
  • E249 adds that “agent native” is an existential product test, not a marketing label: without the agent, the business should lose its core reason to exist.

Connections