Summary
Highlights
The Startup Podcast features Jeremiah Owyang, a leading expert in AI, to discuss the future of AI and agents. The episode introduces concepts like B2A (Business to Agent), meta-agents, and execution agents. Jeremiah, an investor at Blitzscaling Ventures and host of Llama Lounge, defines an AI agent as autonomous software that perceives, makes decisions, takes actions, and learns over time, contrasting it with traditional scripted software. He provides examples such as agents planning meals, managing emails, or collecting work data.
The discussion emphasizes that AI agents differ from traditional software by focusing on achieving goals rather than following explicit instructions. The speakers highlight that agents are non-deterministic and can adapt their methods. They differentiate between simple generative AI and long-lived agents that perform multi-step, adaptive processes, citing examples like Waymo cars as sophisticated agents acting in the real world.
The central thesis is that AI agents will become the primary interface, navigating the internet and applications to gather information and complete tasks on behalf of users. This shift will lead to personalized information delivery and potentially reduce the need for numerous apps. The speakers debate the impact on B2B and B2C SaaS companies, suggesting that many traditional applications might be replaced by dynamically generated user experiences or direct agent-to-agent interactions, especially for mundane tasks.
Jeremiah explains that the data/content layer is decoupling from the presentation layer. Advertisements could be affected, and new business models, potentially subscription-based, will emerge for premium agents. He predicts APIs will evolve, with Anthropic, Google, MIT, and Stanford developing new agent communication protocols like MCP and ADA, aiming for more fluid and dynamic interactions that do not require pre-made integrations.
Salesforce and HubSpot have pivoted towards agent-based strategies, with Salesforce renaming Dreamforce to Agent Force and HubSpot launching an agent marketplace. Skyfire is introduced as an example of an agent handling financial transactions with verified identity and guardrails. Crew AI, an 18-month-old company, is noted for its rapid growth, operating with a small human team and 150 AI agents, highlighting its scalability. Google and OpenAI are also actively developing agentic capabilities.
The most common use cases for agents involve automating repetitive tasks in customer care and reporting. The agent ecosystem is mapped into four layers: data access, management (permissions, orchestration, arbitration, payments, learning), application, and platforms/marketplaces. Opportunities for startups are identified in developing deep vertical agent applications and fostering interoperability between different agent ecosystems. The discussion highlights the trend of agents cannibalizing tasks within roles and industries, leading to the potential for fully agentic companies.
Jeremiah forecasts significant changes, including a 25% drop in search engine volume and the potential emergence of 'autonomous organizations' run by no human employees, offering services and generating revenue independently. The conversation also predicts the evolution of APIs from structured REST APIs to conversational, adaptive agent-mediated protocols, where communication will increasingly be in natural language, making traditional APIs less relevant as tokens become cheaper.