AI Agents Evolve: From Harnesses to Autonomous Claws

Sam Bhagwat, CEO of Mastra, a Typescript agent framework, delivered a presentation titled "Every Harness Will Become A Claw" at the AI Engineer World's Fair. Bhagwat outlined the progression of AI agents, moving from simple Large Language Models (LLMs) to more sophisticated entities capable of complex tasks and autonomous action.

The Agentic Spectrum: From LLM to Claw

Bhagwat positioned the evolution of AI agents on a spectrum, starting with LLMs, which are enhanced by features like tool calls, memory, and retry capabilities to become 'Agents'. These agents further develop into 'Harnesses' by gaining durability, planning modes, and the ability to manage parallel sub-agents.

The next stage, 'Always-On' cloud harnesses, signifies agents that are persistently available and can interact across various platforms like Slack, mobile apps, and cloud sandboxes. This transition enables greater parallelism and more complex workflows, such as PR-native operations.

The ultimate evolution, according to Bhagwat, is the 'Claw', characterized by initiative and learning. These agents can proactively engage, such as by monitoring external feeds for urgent information or learning from their own operational traces to improve performance. He noted that while concepts like automated skill generation are emerging, the industry is still exploring the best methodologies for continuous learning.

The Shift to Cloud Harnesses and the Drive for Initiative

A significant trend highlighted is the move from local to cloud-based harnesses. These cloud-native agents offer enhanced capabilities, including increased parallelism and the ability to tunnel to local machines or operate independently in cloud sandboxes. Bhagwat emphasized that this shift requires different architectural considerations but ultimately leads to more powerful agents.

The transition from harnesses to 'Claws' is marked by the infusion of initiative and learning. Bhagwat described initiative as an agent having a 'heartbeat,' waking up periodically to perform tasks or respond to external triggers. This proactive behavior, combined with continuous learning mechanisms that allow agents to self-improve based on their actions and outcomes, defines the Claw stage.

Steinberger's Law and the Future Shakeout

Bhagwat introduced "Steinberger's Law," a prediction that "Every harness will expand until it becomes a Claw." He attributed this to both technological advancements and psychological factors, noting that users desire more capable and proactive AI assistants. The drive for more functionality, from interacting via direct messages on Slack to initiating overnight tasks, fuels this expansion.

Looking ahead, Bhagwat predicted a significant shakeout in the AI agent market. Drawing a parallel to the consolidation of mobile platforms like Android and iOS, he suggested that users will only have the mental capacity for a limited number of highly effective "Claws." This evolutionary pressure will favor agents that are either highly economically valuable or frequently used, leading to a consolidation where only the most compelling user experiences survive.

For developers, Bhagwat advised ensuring their agents possess the necessary capabilities to meet user needs and to stay abreast of the rapid pace of innovation. He cautioned that without differentiation, agents risk becoming obsolete as more powerful alternatives emerge.