Anthropic Releases 6 Principles for AI Agent Collaboration

On August 19, 2026, Anthropic released six principles for organizational operations where people and AI agents work as a team, through a conversation with Slack's Chief Product Officer (CPO) Jaime DeLanghe. This article covers Slack's practical operational methods, which have aimed to 'transform work conversations into organizational knowledge' since before the emergence of AI.

The core is the multiplayer agent experience through Claude Tag. Moving away from the 'single player' structure where one person works one-on-one with one AI, this approach involves adding Claude as a team member to Slack channels, allowing anyone to mention @Claude to delegate tasks and share results.

From Conversations to Knowledge Slack diagnosed that the promise of past conversation records automatically accumulating into organizational knowledge had not been realized, as humans could not process the vast volume of conversations. Now, by having agents quickly read and analyze massive amounts of text, it has become possible to treat conversation records as a searchable knowledge base rather than simple logs.

Practical Principles and Technical Foundation

• Set Public Channels as Default: Since agents learn only from public context, decisions and discussions must remain in public channels.
• Request for Reasoning: Instead of searching for records of results, ask agents to reconstruct 'why they made that decision' and changes in context.
• Widen Surface Area: Connect Slack, meetings, emails, document repositories, etc., to broaden the context accessible to agents.

MCP (Model Context Protocol) and the Real-Time Search (RTS) API are mentioned as technologies supporting this approach. Slack enabled external agents to access conversation contexts via standard protocols, allowing over 50 partners including Anthropic, Google, and OpenAI to build context-aware agents. The RTS API is designed to query data in real-time reflecting permissions without copying data externally, making it suitable for organizations requiring security reviews.

Additionally, Jaime DeLanghe defined documents as 'receipts' and conversations as 'Thick Work', emphasizing that the conversational context leading up to the creation of a document is more valuable learning material for agents than the refined document itself.