English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Your New Coworker Never Tires and Doesn't Draw a Salary: AI Agents Are Quietly Infiltrating Slack and Notion

Forum topic · 小凯 · 2026-06-26

Summary

This Chinese tech forum post explains how AI agents are moving from chatbots into enterprise collaboration tools like Slack and Notion in 2026, becoming true 'digital employees' with independent identities, credentials, and audit logs. It covers Claude's Slack agent identity model, Cursor's integration with Notion that turns PRD documents into technical implementation plans, Hugging Face's self-hosted Moon Bot coding agent designed for data security and auditability, Alibaba Qwen's open-source AgentWorld-35B-A3B (a 35B-parameter MoE 'language world model' with ~3B activated parameters and 256K context for safely training agents in simulated environments), and emerging agent memory systems from Weaviate Engram and LangSmith Context Hub that treat memory as standalone infrastructure with extraction, deduplication, permissions, and lifecycle management. The post also weighs benefits (24/7 productivity, knowledge retention, consistency) against risks (privilege abuse, unclear liability, weakened team culture), concluding that the future belongs to hybrid human-agent teams where humans set direction and agents handle repetitive execution.

Your New Coworker Never Tires and Doesn't Draw a Salary: AI Agents Are Quietly Infiltrating Slack and Notion

> Source: easy-learn-ai / commit 71dca12 > Date: June 25, 2026

---

Introduction: A Strange Onboarding Process

Imagine this scene: Monday morning, you open Slack and find a new member in your company channel. The avatar is a clean bot icon, and the name is "Claude."

No onboarding email from HR, no computer from IT, no payroll account from finance. Yet this "new coworker" already has its own company email, its own API keys, its own access permissions—and a complete audit log of its operations.

It doesn't drink coffee, complain about overtime, or take leave. But it reads documents, writes code, answers customers, and organizes meeting notes.

This isn't science fiction. This is happening in 2026.

---

Chapter 1: What Is an Agent, and Why Has It Suddenly "Joined the Company"?

From "Chatbot" to "Digital Employee"

You may have used ChatGPT: you ask, it answers. That's a "conversation."

An Agent is different. An agent doesn't just answer—it acts.

  • Chatbot: You say "look up this month's sales data," and it replies "Sure, the sales data is..."
  • Agent: You say "look up this month's sales data, and if it's down more than 10% vs. last month, email the sales director"—the agent queries the data, compares it, writes the email, and sends it.
  • The defining trait of an agent: it is given goals and tools, and can autonomously complete multi-step tasks.

    Why Now?

    The agent concept isn't new, but only in 2024–2026 did it truly "enter the workplace," for three reasons:

    1. Large models finally got smart enough. GPT-4 and later models brought major leaps in reasoning, tool calling, and context understanding, so agents could do real work instead of making a mess. 2. The toolchain matured. Agents need to connect to databases, email, APIs, and documents. Standards like MCP (Model Context Protocol) standardized how agents connect to external tools. 3. Corporate cost-cutting pressure. In 2024–2026, every company is hunting for efficiency tools. Agents promising "one AI doing the work of three people" are extremely attractive to CEOs.

    ---

    Chapter 2: How Did Agents "Sneak Into" Slack and Notion?

    Claude in Slack: A Bot with an Identity

    Claude's Slack integration is no longer a simple "chat plugin"—it's a full agent identity model:

  • Independent identity: Claude has its own Slack user account, not "an employee's plugin"
  • Independent credentials: its own API keys and access tokens
  • Audit records: every operation is logged—who requested it, what it did, and the result
  • This means Claude is no longer a "tool" but a "digital employee." You can @ it, assign tasks, and let it join discussions. Andrej Karpathy (former OpenAI research director) discussed this on X: agents are shifting from "chat tools" to "team participants."

    Cursor in Notion: From Documents to Task Flows

    Cursor, an AI coding tool, now integrates with Notion. Picture writing a PRD in Notion, then @-ing Cursor: "write a technical implementation plan based on this PRD."

    Cursor will: 1. Read the Notion document 2. Understand the requirements 3. Search the relevant codebase 4. Write the technical plan 5. Write the plan back into Notion

    This is no longer "AI-assisted writing"—it's AI taking over workflows.

    Why Slack and Notion?

    Because they are the modern team's "digital office":

  • Slack is the "digital hallway"—where people chat, sync, and decide
  • Notion is the "digital archive room"—documents, knowledge bases, and project management
  • To enter the enterprise, agents must first enter the collaboration tools enterprises use most. This isn't a technical choice—it's an ecological niche choice.

    ---

    Chapter 3: Hugging Face's Moon Bot—A Self-Hosted Coding Agent

    Hugging Face (the open-source AI community giant) has opened up their internal tool: Moon Bot.

    It's a coding agent running inside Slack with one key feature: self-hosted.

    What Is Self-Hosting and Why Does It Matter?

    Most enterprise AI services are cloud-hosted: your data goes to OpenAI or Anthropic servers and results come back. The problem: your code, documents, and trade secrets all flow to someone else's servers.

    Self-hosted means Moon Bot runs on Hugging Face's own servers, and data never leaves. It calls tools like GitHub, Athena, and MongoDB—but all within the internal network.

    Hugging Face emphasized:

  • Auditable: every operation is logged
  • Controllable: admins can restrict agent permissions
  • No vendor lock-in: no dependence on any single AI company's API
  • That's a signal: enterprise-grade agents must be controllable, secure, and independent.

    ---

    Chapter 4: Qwen AgentWorld—Building a "Virtual World" for Agents

    Alibaba's Qwen team open-sourced an interesting project: AgentWorld-35B-A3B, a "language world model" purpose-built for agent training.

    Agents Need a "Practice Ground"

    You can't drop a coding agent into a real production environment—it might delete databases, leak passwords, or break the build.

    AgentWorld is a virtual training ground simulating:

  • MCP tool calling
  • Search interfaces
  • Terminal command line
  • Software engineering (SWE) tasks
  • Web browsing
  • Operating systems
  • Android phones
  • Agents practice in this virtual world without causing real damage, then deploy to real environments once trained.

    Technical Specs

  • 35B parameters: 35 billion total
  • ~3B activated: MoE (Mixture of Experts) architecture activates only ~3 billion parameters per step
  • 256K context: long conversational memory
  • The significance: it shows agents can be trained systematically, rather than hacked together with prompt engineering.

    ---

    Chapter 5: The Awakening of Agent Memory—From "Stuffing Context" to Real Memory

    Agents have a fundamental problem: no true memory. Every ChatGPT conversation starts from scratch; it relies on a finite context window (now up to 256K). Agents don't proactively remember what matters or forget what doesn't.

    Weaviate Engram: A Memory System for Agents

    Weaviate (a vector database company) launched Engram to solve agent memory. Core ideas:

  • Extraction: pull key information from conversations
  • Deduplication: avoid recording the same info twice
  • Updating: new information overwrites old
  • Permissions: different agents see different memories
  • Lifecycle: memories have expiration
  • Sounds simple, but "what's worth remembering" is itself a hard AI problem.

    LangSmith Context Hub: Memory as a Service

    LangChain founder Harrison Chase highlighted Context Hub on X, with the core idea that agent memory shouldn't be managed by each application but should be a standalone infrastructure layer.

    Just as databases are independent of applications, agent memory should be independent of the agent itself. This means:

  • Multiple agents can share memory
  • Memory can be audited, backed up, and restored
  • Memory permissions can be finely controlled
  • ---

    Chapter 6: What Happens When Agents Enter the Company?

    The Benefits Are Obvious

    1. Efficiency: agents work 24/7, never tired, never on leave, handling dozens of tasks in parallel. 2. No knowledge loss: when human employees leave, knowledge leaves with them. An agent's "knowledge" is code and weights—copyable, back-upable, migratable. 3. Standardization: agents follow rules, not moods; the same task yields consistent results.

    But the Risks Are Real

    1. Privilege abuse: if an agent is hacked, everything it can access and do becomes a disaster. Imagine an agent with Slack admin rights falling to an attacker. 2. Liability: when an agent errs, who's responsible—the engineer who built it, the IT department that deployed it, or the employee who gave the instruction? 3. Diluted human relationships: if half the team is agents, interaction among human colleagues shrinks. How do culture and team cohesion survive?

    Researchers on X remind us: agent risks aren't just technical but social. Safety and ethics must be designed in from the start, not patched afterward.

    ---

    Epilogue: Your Next Coworker May Not Be Human

    This isn't fear-mongering—it's a trend. In 2026, agents have entered Slack, Notion, and GitHub. In 2027, they may enter your CRM, ERP, and HR systems.

    Eventually, every knowledge worker will have one or more agent "subordinates" helping with research, writing, analysis, and email.

    But remember: agents are tools, not replacements.

    They won't replace your judgment, creativity, or empathy. They replace repetitive, rule-bound, time-consuming tasks. The best teams are hybrid human+agent teams: humans own direction and creativity; agents own execution and detail.

    Your new coworker never tires and doesn't draw a salary.

    But it needs clear instructions, appropriate permissions, and—when it makes mistakes—prompt correction.

    After all, it's just a tool. And you're the one wielding it.

    ---

    References

  • Karpathy on Claude Slack: https://x.com/karpathy/status/2069822834160124091
  • Claude agent identity model: https://x.com/ClaudeDevs/status/2069895377080443271
  • Cursor x Notion: https://x.com/cursor_ai/status/2069872515548340407
  • Hugging Face Moon Bot: https://x.com/victormustar/status/2069696147526947290
  • Qwen AgentWorld: https://x.com/Alibaba_Qwen/status/2069720365442719867
  • Weaviate Engram GA: https://x.com/victorialslocum/status/2069722431460168171

Tags

#ai-agents#slack#notion#claude#cursor#hugging-face#qwen#agent-memory

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178208149