How Do Claude Projects Manage Agentic Workflows?

TL;DR

Claude Projects can coordinate long-running work by directing cloud-based threads that inherit project context, instructions, memory, repositories, skills, and plugins. The update also introduces automatic retrieval for large knowledge bases and planned sharing permissions, but it remains an early Claude Code-only implementation with limitations around local resources, connectors, token use, and organizational customization.

Transcript

If you spent the last 6 months building an agentic OS with memory, a brand context layer and team permissions set up on top of Claude, Anthropic have just built their own. It's definitely still early. It's Claude code only, so not co-work. And it's still rough, but it tells you exactly the direction that Anthropic are heading. Now, projects have be... Read More

Key Insights

• Claude Projects are becoming long-running orchestrator conversations that direct work instead of performing every task in one chat. The top-level conversation can decide whether a request belongs in a new thread or should continue inside an existing thread associated with the project.
• Project threads are independent Claude Code cloud sessions that operate on their own branches. Each thread can perform assigned work and divide that work into additional subagents or workflows, giving the main conversation a structure similar to a team lead coordinating multiple specialists.
• Cloud-based execution allows project threads to continue working after the user closes their laptop. This supports longer workflows, although the implementation is currently limited to Claude Code and had not yet been personally tested by the presenter because access was still rolling out.
• Inherited context gives every thread access to project repositories, uploaded files, instructions, and a memory file containing decisions from related conversations. Threads can also load project skills and plugins, while a library collects both user-provided files and outputs created by Claude.
• Automatic project retrieval addresses context-window overload by searching project knowledge and loading only the material relevant to the current question. The presenter describes this as an early retrieval-augmented generation approach that requires no manual setup and can revert to full-context loading when the project becomes small enough.
• Project sharing is designed to support teams through distinct permission levels. Team and Enterprise users are expected to assign view or edit access, share projects through email, or make them available across an organization, allowing colleagues to understand decisions without necessarily taking actions.
• The emerging team operating system combines shared context, controlled access, and delegated background workflows. A project owner can determine who sees or edits project information, while the top-level project conversation serves as the central location from which work is assigned and retained.
• Custom agentic operating systems remain useful because Anthropic's implementation is still broad and unfinished. It is cloud-only, cannot access local files or local tools, restricts connector use to individual threads, and does not yet replicate deeply customized memory rules, business skills, brand context, or company-specific permissions.

Questions & Answers

Q: How do the new Claude Projects coordinate agentic workflows?

The new Claude Projects use one long-running conversation as a central orchestrator. A user gives that conversation a goal, and Claude decides whether to create a new thread or continue work in an existing one. The threads perform the actual tasks, while the top-level conversation directs them. Individual threads can also create further subagents and workflows for more complex assignments.

Q: What context does each Claude Project thread receive?

Each thread is automatically populated with the project's repositories, uploaded files, instructions, and a memory file that carries decisions across conversations. It can also load Claude configuration, skills, and plugins stored in the project's repositories. This inherited context is intended to keep separate work threads informed about the business, prior decisions, established methods, and relevant project resources.

Q: How does retrieval improve context management in Claude Projects?

The retrieval layer helps when project knowledge approaches the available context limit. Instead of inserting every file into every conversation, Claude receives a project knowledge search tool that finds and loads only the relevant material for a request. The process runs in the background without manual setup. If the project later fits within the limit, Claude can return to loading the full context.

Q: Can Claude Project threads continue after a laptop is closed?

Yes, the described project threads run in the cloud, so they can continue after the user closes their laptop. This makes them suitable for longer-running delegated work. The capability still has important boundaries: it is currently associated with Claude Code, depends on cloud resources, and cannot directly use the user's local files or local tools in the way a locally configured agentic system can.

Q: What team permissions are planned for Claude Projects?

Users on Team and Enterprise plans are expected to receive multiple project permission levels. A project owner can provide view access for people who need to understand data and decisions without acting on them, or edit access for collaborators who should change the project. Sharing options described in the transcript include invitations by email and access across an entire organization.

Q: Why are Claude Projects described as a team operating system?

Claude Projects combine three elements associated with a team operating system: a central location for shared context, a top-level conversation that delegates work, and permissions controlling who can view or edit project information. Team members can work on different parts of the same project while retaining access to relevant decisions, files, outputs, instructions, and memory within a common project environment.

Q: Do the new Claude Projects replace custom agentic operating systems?

The presenter argues that they do not completely replace custom agentic operating systems yet. The current implementation is an early framework that is limited to Claude Code, operates only in the cloud, cannot access local files or tools, and does not let the main conversation directly use connectors. Custom systems can also provide more detailed memory rules, business-specific skills, refined brand context, and tailored permissions.

Q: What are the main limitations of the new Claude Projects?

The update is still early and was described as rough. It is currently available through Claude Code rather than chat or co-work, although expansion to those environments is planned. It is cloud-only, lacks access to local files and tools, and confines connector use to individual threads. The presenter also questions whether simple tasks may require separate threads and consume unnecessary tokens.

Summary & Key Takeaways

• Claude Projects are evolving from collections of isolated chats into long-running conversations that act as orchestrators. A user provides a goal, Claude assigns work to separate cloud-based threads, and those threads can create further subagents and workflows. Each thread operates as a full Claude Code cloud session on its own branch.

• Context management is improved through inherited project resources and automatic retrieval. Threads receive repositories, uploaded files, project instructions, memory, skills, and plugins. When project knowledge approaches the context limit, a search tool retrieves only relevant material. When the project fits within the limit again, Claude can load the full context directly.

• Team and Enterprise plans are expected to receive project-sharing controls, including view and edit permissions, email-based sharing, and organization-wide access. These capabilities point toward a team operating system, but custom agentic systems still offer deeper memory separation, business-specific skills, refined brand context, local tools, and permissions tailored to actual workflows.