Video summary

Anthropic Just Built It’s Own Agentic OS

Main summary

Key takeaways

Technology

Overview

Anthropic introduced a new “agentic OS” style feature set inside Claude—positioned as a next-generation version of Projects. It focuses on:

  • Better context handling
  • Agent/thread delegation

Note: It’s still early and rough. It currently works code-only (Claude Code / projects stored there) and is cloud-only, with a phased rollout.

What’s new: Projects become “one long-running conversation” with delegation

Previously, Projects in Claude chat/co-work were mainly organized contexts where you could open separate chats/sessions within a project.

That approach often breaks for large tasks because you can hit context-window limits, forcing you to split into new sessions—losing continuity.

With the update:

  • A Project is now treated as an ongoing conversation
  • It delegates work instead of requiring you to manually open/coordinate many separate chats

Thread model (main orchestrator → subagents/workflows)

The system uses:

  • A top-level “main agent” that delegates tasks into threads
  • Threads can spawn:
    • Subthreads
    • Subagents
    • Workflows

In plain terms:

  • Threads do the work
  • Claude directs the work (like a team lead)
  • Claude decides whether to reuse an existing thread or start a new one
  • Threads run in the cloud and can continue after you close your laptop

Context persistence and shared business knowledge inside threads

Each thread is automatically populated with:

  • The project’s uploaded files/repos
  • Project instructions
  • A memory file (mentioned as memory.mmd) carrying decisions across conversations
  • The project’s Claude code repos, skills, and plugins, helping preserve:
    • brand voice
    • assets
    • prior tooling
  • A library tab that aggregates:
    • files you added
    • files Claude produced

This is intended to prevent outputs from getting lost across many isolated sessions.

Key improvement #1: Retrieval layer for context (RAG-style)

Projects now include a retrieval layer described as RAG (retrieval-augmented generation).

Instead of loading everything into each conversation, the system:

  • Searches the project’s knowledge
  • Loads only the relevant parts for the current question
  • Shrinks back under the context limit when less context is needed

Key points:

  • Intended to be automatic (no user configuration)
  • The creator notes similarity to RAG-style “OS memory systems,” but frames this as an MVP with zero setup cost
  • A separate “RAG guide for projects” is referenced

Key improvement #2: Team collaboration with permissions

For team/enterprise plans, Projects support permissions:

  • View-only users: understand decisions but can’t act
  • Edit users: can modify work

Sharing options include:

  • By email
  • Organization-wide sharing

Goal: make collaboration work even though tasks may happen across isolated environments/threads, by ensuring shared access to project context and decision history according to permissions.

Availability + limitations (important caveats)

  • Currently available only to Claude Pro and Max (rolling out); the narrator does not have access yet
  • Not available for Chat / co-work—it’s Claude Code only
  • Cloud-only: cannot access your local files/tools
  • Mentioned limitation: the main conversation can’t directly touch connectors like MCPs; those are handled independently in threads

Takeaway from the analysis

The video argues Anthropic is moving toward an “agentic OS” direction via:

  • context injection
  • delegation
  • memory
  • skills/brand context
  • permissions

However, the creator suggests it doesn’t fully replace agent systems yet because:

  • it’s not fully integrated into chat/co-work yet
  • it’s cloud-only with connector limitations
  • it’s still an MVP framework and missing deeper “OS” layers (e.g., richer personalization vs shared data, more advanced skill/brand systems, etc.)

They expect these frameworks to evolve—and potentially be replaced by other providers later—so they plan to migrate as needed.

Main speakers / sources

  • Primary speaker: the YouTube narrator/reviewer discussing Anthropic’s announcement and their own “agentic OS”/community tools
  • Source material referenced: Anthropic’s official release/video explaining the updated Projects feature

Original video