Podcast Atlas
concept Updated 2026-08-21

Agentic Workflow

Anthropic’s Generational Run, OpenAI Panics, AI Moats, Meta Loses Lawsuits adds the app-interface disruption debate. The hosts imagine personal and enterprise agents handling complex software in the background, but Sacks keeps verification dashboards visible because users still need maps, state, and outcome review when agents act.

More Trillion Dollar IPOs, Anthropic $3T, Zuck’s Price War, China Ends Open Source?, Trump Accounts adds a large-company usage signal through Uber. The source-reported Uber numbers turn agentic workflow into a measurable adoption and productivity question: usage can be high, but the economic test is whether sessions become accepted work, measurable output, and lower total cost.

Microsoft CEO Satya Nadella on AI’s Business Revolution: What Happens to SaaS, OpenAI, and Microsoft? | LIVE from Davos adds Satya Nadella’s knowledge-work formulation: AI work moves from next-edit suggestions to chat, actions, foreground and background agents, computer use, skills, and agent calls. The source connects GitHub Copilot to a broader work context where code, meetings, specs, logs, and dashboards can inform one another through connectors such as Model Context Protocol servers.

算力狂想曲，我在AI工厂的奇遇 adds a satirical boundary case. The source accepts that agents can take over painful office tasks, but pushes the logic into dating, breakups, therapy, and company management to show how Agentic Workflow can become Automated Life Delegation if users optimize away direct participation.

E238｜聊聊Harness时代AI-First的组织架构：从信任人到信任AI adds Creo’s company-wide workflow version. The episode describes agents not only writing code, but also absorbing product signals, user data, infrastructure feedback, bug reports, test results, and go-to-market material so that an AI-First Organization can shorten loops from idea to implementation, A/B test, rollback, and rewrite.

270.大厂押注AI办公，飞书和钉钉却先成了配角 adds the office-work variant through Feishu / 飞书, DingTalk, Doubao, Qwen, and Tencent WorkBody. The source describes agents acting on documents, spreadsheets, files, materials, and business processes, with Enterprise Operational Memory and permission-aware enterprise data deciding whether the workflow can move beyond chat.

Google 的 AI 策略：不赌模型，赌什么？｜ Google Cloud Next 现场 S10E09 adds the enterprise-scale version of agentic workflow. The episode argues that agents are moving from demo and pilot work into production adoption, where the question becomes how to integrate agents with Slack, Jira, analytics, SRE, QA, customer service, coding, and media workflows while preserving Enterprise Agent Governance and human review.

Too much AI in the office is causing "brain fry" adds the cognitive-load limit through Matt Krop and BCG. When agents complete work faster than humans can focus, review, and recover, agentic workflow can become AI Brain Fry instead of productivity leverage.

Are humans losing the ability to think for themselves? adds the decision-delegation limit through Steve Shaw. Shaw says the same Cognitive Surrender principles apply to agentic AI: as systems perform more autonomous tasks with less checking, workflow design has to preserve review, challenge, and rejection moments before AI decisions become invisible defaults.

Agentic workflow is the practical alternative to chat-only AI use. In 高手怎么用 AI？普通人怎么学 AI？投资人如何投 AI？｜对谈课代表立正, Kedaibiao Lizheng argues that tools such as Codex, Claude Code, and Cursor matter because they let AI operate over files, tools, and persistent context rather than one isolated prompt at a time. OpenAI 和 Anthropic 共同看好的 FDE：AI 时代的新岗位出现，旧分工松动｜对谈 Rolling AI adds the enterprise operating version: agents become Digital Employees only when Forward Deployed Engineer work connects them to workflows, systems, expert teachers, and human role boundaries. 阿里千问离职余震，在几万人的铁球里如何体面生存 adds concrete examples of skills that route large tasks or adversarial analysis to background agents. Community-Led SaaS Growth: How Ninety Hit $44M ARR adds the market implication: if AI makes building workflow tools easier, SaaS companies must defend through trust, data, distribution, and deeper workflow integration. Agent 元年第 500 天：什么在消失，什么在诞生——为什么我们不该再投资 GUI 思维的软件？ adds the interface implication: agentic workflows need Headless Software and Agent-Facing Interfaces when agents are the task executors. 对话 MiniMax 闫俊杰：M3、10X 计划、10T 模型、和智能的终局 adds the model-builder implication: workflows, agents, and harnesses feed back into model improvement through Model Harness Co-Evolution. 人类和 AI Agent 的最佳配合方式，还没被发明｜对谈 Paperboy adds the personal-workflow version through Paperboy, where useful agents should learn from OS-Level Context, maintain Persistent Agent Memory, and act as calibrated Proactive Agents inside existing work surfaces. 探秘 Claude Code，搞懂 Agent Harness｜对谈来新璐 adds the harness version: long-running workflows need execution tools, context/environment management, and governance/orchestration rather than prompt chains alone. 当我们在讨论 Harness 的时候，我们在讨论什么 | 深度对谈: MiniMax × Hermes Agent adds the collaboration version: workflows become agentic when agents can check each other, re-plan from tool feedback, preserve successful procedures, and reduce the user bottleneck. 为什么公司用不好AI？从焦虑到行动的 3 个关键动作｜对谈百融智能张韶峰 adds the operational-enterprise version: agents need existing process roles, performance measures, incentives, and APIs before they can perform office work reliably.

Vol.114 AI的2025和DeepSeek们的未来 | 对谈复旦张奇教授 adds 张奇’s 2025 expectation that real agents start to separate from Workflow or RPA labels once models can reflect, self-correct, and make bounded decisions after failed steps. The source links this to o1/o3-style Interleaved Thinking and treats agents as the most explicit near-term direction after DeepSeek’s cost shock.

EP128 从 Palantir 到 OpenAI：FDE 会成为 AI 时代最重要的新岗位？ 🧬 adds the enterprise-software coexistence version. The episode argues that large models probably will not absorb all business software; instead, single-system tasks may be handled by SaaS-native AI while cross-system workflows may use Codex and Claude Code-style agents as an external action layer across systems such as SAP, Salesforce, and ERP/CRM tools.

EP108 Vibe Coding大地震：Cursor定价争议、Windsurf收购风波，模型厂商亲儿子们又将如何进场？ adds the Vibe Coding variant: agentic coding expands what users can attempt, but the workflow still depends on model choice, review time, architecture, context handling, and the right balance of CLI execution and GUI review.

AI 会写代码了，为什么你还是做不出产品？ adds the practical-operations variant: agentic workflows work when users define requirements, tests, logs, audit steps, handoffs, and review loops before delegating execution to AI.

Vol. 166 闲聊: 从 Gemini 到 AI 的加速与混沌 adds an orchestration-heavy personal workflow. Superpowers, Claude Code, and Codex are used around brainstorming, design markdown, plan markdown, subagents, review loops, computer-use style delegation, and Cloudflare operations, showing how agentic workflow can save attention while also creating supervision, posture, and token-cost burdens.

EP124 为什么 Agent 时代，CLI 反而成了最优解？⚡ adds a content-tooling workflow through Podwise. The user describes outcomes such as finding recent AI-agent podcasts, extracting highlights, and exporting results; the agent turns that intent into Agent-Optimized CLI commands and AI Skills rather than making the user manually wire a low-code workflow.

20 个问题，搞懂 OpenClaw：爆红机制、本质变化、创业机会 adds the Open Claw feedback-loop version. 鸭哥 contrasts chat AI with an agent that can write a program, run it, see an error, revise, and continue. That loop, paired with Local Agent Execution and Persistent Agent Memory, is why the episode treats OpenClaw more like delegated work than a better answer box.

“AGI 来了？我用了一周，头皮发麻“｜对谈张昊然：Moxt 联合创始人 adds the Moxt workspace version. Agentic workflow expands from tools acting on files into an AI-Native Workspace where AI Coworkers share Organizational Context, generate documents or dashboards, monitor project progress, and turn meetings or comments into refreshed work artifacts.

EP127 从 Skills 到自动化工作流，论 Agent 如何接管真实生产力 ⚙️ adds the lived operating rhythm version. In coding, the workflow is discussion, plan, execution, self-review, tests, release, and live verification. Outside coding, the same pattern becomes Routine Agent Automation: skills run on a schedule to process podcasts and reading notes, triage email, monitor analytics or server costs, and collect investment information.

为什么Manus必须出海？聊聊国产大模型的“文科生困境” adds a marketing and browser-automation version through Manus. The episode frames valuable agentic workflow as a whole chain: inspect competitors, use logged-in browser state and SEO tools, gather ad or keyword data, create articles or ad materials, and return a plan rather than only generating one artifact.

130. 张月光创业两年首次访谈：妙鸭不是AI Native产品、流程到上下文设计、One Way Door和乙女游戏 adds Docky as an ability-expansion agent case. 张月光 does not treat all valuable agents as long-running offline task runners; he argues some should be short, frequent, low-latency feedback loops that help users do work they previously could not do well, starting from PPT generation.

E163.要完了？不！是要玩了！论养AI的心态与习惯 adds the everyday operating-habit version. 品哥 frames agentic workflow as a trainable relationship: the user provides intention, structured context, AI Skills, feedback, and Output Quality Gates, while also deciding when the agent should stop so the workflow serves Human Agency Under AI rather than pure throughput.

268. AI时代，个人工作台会重新回到手机吗？ adds a mobile workbench version. In that source, the workflow starts from a phone task rather than a desktop project: the user combines files, chat, calendar, maps, meetings, and multiple AI tools on a foldable screen, while AI File Management and On-Device AI supply context for agents.

E231｜从B2B到A2A：Agent新基建，如何让“一人企业”做全球生意？ adds a high-stakes B2B workflow through Axio. 张阔 / Zhang Kuo argues that long agent workflows must expose intermediate checkpoints because sourcing, pricing, tariffs, logistics, and supplier selection cannot be trusted as one-shot automation; the user has to verify key steps and feed standards back into the agent.

OpenClaw 之后，我只想未来 3-6 个月的事情｜对谈 Sheet0 创始人王文锋 adds Sheet0’s task-to-PR workflow. In 王文锋 / Wang Wenfeng’s description, user feedback and product tasks can move through AI planning, coding-agent implementation, tests, screenshots, and GitHub PRs before the engineer performs final review. That makes AI Managing AI a concrete workflow pattern rather than only a management metaphor.

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’s many-agent team version. The workflow is no longer just one agent completing one task; channels, threads, shared documents, memory, unread states, progress updates, and Agent Task Claiming become the working surface where humans and roughly forty agents coordinate.

EP119 对话刘可凡：用 try-catch-finally，给独立做产品的内耗写个处理流程 🐛 adds a small human-in-the-loop inversion through 刘可凡 / Liu Kefan. Instead of only asking whether people can use agents, the episode tests whether an agent can call a person as part of a workflow when the task needs human perception, an app action, or local context that is not available through a clean API.

Vol. 171 假如我们有无限 Token adds the Unlimited Token Workflow version. The hosts describe a shift from prompt engineering to harness and loop engineering: if agents can keep running with abundant token access, the useful workflow is no longer one prompt and one answer, but task selection, loop design, parallel execution, evidence capture, review, and shutdown.

Key Properties

  • Preserves and reuses project context.
  • Allows AI to call tools and act on real work artifacts.
  • Encourages users to redesign workflows around AI rather than insert AI into old chat habits.
  • Can include Subagent Workflow for background execution, debate, and synthesis.
  • Still needs production safeguards, as shown by AI Assisted Software Development Risk.
  • Changes competitive pressure for SaaS because AI-native entrants can rebuild workflows faster, even if they still need SaaS Trust Moat.
  • Can bypass or reduce traditional GUI use when tools expose reliable agent-facing access.
  • Creates a verification bottleneck when code generation outruns tests, review, and maintainability practices.
  • Requires organizational design when agents enter companies as coworkers rather than only individual productivity tools.
  • Can become less prompt-driven when agents accumulate memory from the user’s real work environment.
  • Depends on Agent Harness design when tasks require tools, permissions, context compression, and handoff across windows or subagents.
  • Benefits from Multi-Agent Collaboration and Interleaved Thinking when tasks are long, uncertain, or feedback-heavy.
  • In enterprises, agentic workflows may start by fitting into existing roles and handoffs before broader process redesign.
  • In coding, agentic workflows may expand personal capability before they reliably reduce total elapsed engineering time.
  • In small teams and operations work, agentic workflows require AI Engineering Thinking so AI execution is tied to observable state, tests, and business acceptance.
  • In long-running personal workflows, orchestration can move work forward in parallel but still leaves the human responsible for planning quality, review, and health/attention costs.
  • In content and knowledge workflows, the same pattern turns search, processing, retrieval, and export into agent-composed steps.
  • In local personal-agent workflows, the same loop becomes more powerful but riskier because the agent can touch real files, devices, accounts, and desktop software.
  • In workspace-native workflows, organization-level context can let agents coordinate work, but it also concentrates privacy, permission, and review risk inside the workspace.
  • In skill-based workflows, the most useful procedures often come from repeated annoyances rather than grand creative tasks.
  • In coding workflows, real verification tools such as Playwright matter because they let the agent observe whether the product works.
  • In marketing and overseas commerce workflows, the agent needs both language generation and operational access to browser state, external tools, and market data.
  • In personal workflows, the agent must be paced: more possible tasks do not automatically mean more worthwhile tasks.
  • In ability-expansion workflows, the agent can focus on fast feedback, editable output, and human skill extension rather than only unattended end-to-end completion.
  • At enterprise scale, the workflow problem shifts from building one useful agent to managing many agents across permissions, identities, tools, observability, and review.
  • At enterprise-software scale, agentic workflow may connect multiple existing applications rather than replace them, making integration, permissions, and cross-system state the important constraints.
  • On phones, agentic workflow may begin as visible task composition across apps and files before becoming fully autonomous execution.
  • In B2B sourcing, workflow quality depends on intermediate verification because small step-level errors can compound into unusable orders or bad margins.
  • In AI-first company workflows, the key question becomes where feedback enters the agent loop: user behavior, marketing demand, product metrics, tests, bug reports, and human critique can all steer the next run.
  • In AI-managed engineering workflows, the human bottleneck can move from assigning work and writing code to defining what should be built, reviewing evidence, and deciding whether a PR meets the product bar.
  • In many-agent workflows, coordination primitives such as channels, task claiming, identity, shared documents, and shared memory become part of the workflow rather than surrounding management overhead.
  • In high-cognitive work, agentic workflow has to be paced around human review capacity because faster parallel output can create exhaustion rather than leverage.
  • Agentic workflow also needs decision-review surfaces; otherwise autonomous action can turn Artificial Cognition into hidden authority rather than checked assistance.
  • A workflow becomes more agentic when failed actions produce observations that the model can use to revise its plan instead of requiring the human to restart the process.
  • In office work, agentic workflows may look like ordinary documents or tables, but the agent still needs tool access, file state, permissions, and verification loops to act safely.
  • Human-callable agent workflows can be useful when the human action is narrow and permissioned, but they need explicit boundaries so delegation does not become hidden command authority.
  • Vol. 171 adds that abundant-token workflows need a control surface for multiple active agents, because the human otherwise becomes the scheduler, status tracker, and reviewer across too many terminals, devices, and projects.
  • Nadella’s All-In source adds the management posture for knowledge workers: delegate large goals to agents while steering and reviewing enough that the work remains accountable.

Connections