Instead of asking which AI agent is best, teams should evaluate agents based on integration, data access, ecosystem growth, and composability to effectively match tools to specific workflows. By focusing on infrastructure and layering specialized agents according to their strengths, organizations can maximize productivity and avoid inefficiencies from forcing one tool to do everything.
The rapid emergence of AI agents has led to a flood of new products, making it challenging for teams to decide which to focus on. Instead of asking which agent is best, the speaker proposes a five-question filter to evaluate agent launches: Does it integrate with existing tools? Can other agents build on it? Does it access important data? Is there a growing ecosystem? And can agents be stacked on top of it? This framework helps cut through the noise by prioritizing infrastructure and ecosystem over flashy demos or benchmark scores, guiding teams to invest their time wisely.
Applying this filter to recent launches reveals distinct use cases. OpenAI’s ChatGPT Workspace Agents excel in managing shared, repeatable workflows across tools like Slack and ChatGPT, making them ideal for team-based tasks that span multiple platforms. Salesforce’s Headless 360, though less flashy, is a critical infrastructure play that exposes Salesforce capabilities via APIs, enabling agents to work directly with CRM data and workflows. This approach positions Salesforce as a foundational layer in the agent economy, allowing various agents to operate within its ecosystem.
Microsoft’s Copilot Wave 3 demonstrates another infrastructure shift by embedding agent capabilities deeply within the Microsoft 365 environment. It leverages Work IQ to access comprehensive organizational data, making it powerful for enterprises heavily invested in Microsoft tools. However, its closed ecosystem limits openness to external agents, making it less suitable for cross-platform or coding-intensive workflows. Meanwhile, Moonshot’s Kimmy K 2.6 offers an open-weight, multimodal agent model with advanced capabilities like swarm orchestration, appealing primarily to developer teams who can self-host and build custom agent infrastructure rather than typical enterprise users.
Perplexity’s Personal Computer on Mac addresses research-heavy workflows by combining local file access, browsing, and multi-step task orchestration, making it well-suited for tasks like market research or competitive intelligence that culminate in polished deliverables. However, it lacks the governance and repeatability needed for team-shared processes native to platforms like Microsoft 365 or Salesforce. Overall, each agent product fits specific workflow shapes, and forcing one tool to cover all tasks leads to inefficiency and wasted resources.
The key takeaway is that the AI agent landscape is evolving into a layered ecosystem rather than a single dominant product. Teams should avoid switching agents casually and instead layer tools based on the nature of the work, data access, and integration needs. The model quality is important but secondary to the surrounding infrastructure, permissions, connectors, and ecosystem. Mastering the judgment to route work to the right agent layer is the new literacy in the AI era, enabling teams to compound productivity gains by leveraging the strengths of multiple specialized agents rather than chasing the latest flashy launch.