Asana Emphasizes Shared Memory for Effective AI Integration

Shared memory and context are becoming essential components for successful AI integration within enterprises, according to Arnab Bose, Chief Product Officer of Asana. Speaking at a recent event in San Francisco, Bose emphasized that providing AI agents with detailed historical context enables them to function effectively as active teammates rather than passive tools. This approach reduces the need for users to repeatedly provide information about business operations each time they assign a task.

In line with this vision, Asana introduced its AI Teammates last year, designed to seamlessly integrate into teams and projects. This integration allows AI agents to inherit sharing permissions and access project histories, ensuring they are well-informed about completed tasks and outstanding issues. Users can select from 12 pre-built agents tailored for common scenarios, such as IT ticket deflection, or create custom agents based on specific needs. Agents can also utilize third-party resources like Microsoft 365 and Google Drive to enhance their functionality.

When an AI agent is created, it operates as a teammate within the project framework. Bose explained that everything done by both humans and AI is tracked, enabling a transparent system that facilitates explainability. To maintain oversight, workflows include checkpoints, allowing human users to provide feedback and request modifications to project elements or research strategies.

The potential for AI integration does not come without challenges. Asana users must navigate an OAuth process to grant access to Claude, an AI model from Anthropic. Ensuring that all knowledge workers understand the integration’s capabilities and the associated OAuth grants remains a significant hurdle. Bose suggested that identity providers could centralize some challenges around authorization, potentially creating a universal directory of approved AI agents and their skills.

Currently, no standardized protocol exists for shared knowledge and memory, according to Bose. His team has received numerous inquiries from partners interested in having their agents interact with the Asana work graph. However, the absence of a common standard complicates these interactions, often requiring bespoke solutions.

Bose identified three pressing questions in AI orchestration: How to build and manage an authoritative list of approved AI agents, how to facilitate secure app-to-app integrations, and how to achieve a unified outcome across multiple platforms. Presently, agent-to-agent interactions tend to be isolated, with different applications like Asana, Figma, or Slack operating independently.

The introduction of the Modern Context Protocol (MCP) by Anthropic offers a promising step forward. This open standard allows AI agents to connect to external systems with a single action, reducing the need for custom integrations for every pairing. However, Bose expressed skepticism about the existence of a perfect solution at this time, underscoring the ongoing need for innovation in this rapidly evolving field.