Every Harness Will Become A Claw — Sam Bhagwat, Mastra

AI Engineer

2,241 views • 22 hours ago

Video Summary

The speaker discusses the evolution of AI agents, moving from basic LLMs to more sophisticated "harnesses" and eventually "claws." This progression involves increased durability, the ability to run continuously, and the introduction of initiative and learning capabilities. The speaker posits that as these agents become more powerful and integrated into daily life, a market shakeout will occur, similar to the mobile app landscape, where only a few dominant "claws" will remain.

The talk outlines the transition through several stages: agents with tool calls and memory, harnesses with enhanced durability and continuous operation, and finally, claws with initiative and learning. The speaker emphasizes the psychological and economic drivers behind this expansion, likening the desire for these agents to a "dopamine casino." The presentation concludes with advice for builders to focus on user needs and anticipate future waves of innovation and consolidation in the AI agent market.

Short Highlights

AI agents are evolving from basic LLMs to more capable "harnesses" and "claws."
"Harnesses" offer durability, continuous operation, and expanded affordances.
"Claws" are characterized by initiative, learning, and proactive behavior.
A future market shakeout is predicted, with only a few dominant "claws" surviving.
Builders are advised to focus on user needs and anticipate ongoing innovation.

Key Details

The Agentic Spectrum [02:06]

The evolution of AI agents is compared to the spectrum of self-driving car autonomy.
Key differences between LLMs and agents include the agent loop, tool calls, memory, retries, and state management.
The spectrum ranges from LLMs to agents, harnesses, and finally, claws.

"there are various aspects of the agentic spectrum between LLMs agents harnesses and claws"

From Agents to Harnesses [03:12]

Moving from agents to harnesses introduces qualities like durability and "doggedness."
Harnesses can run for hours or days, persist streams, and resume tasks.
Features include planning mode, parallel sub-agents, TUI/slash commands, dynamic agent creation, and background bash tasks.

"durability and doggedness"

The Always-On Cloud Harness [05:09]

A significant shift is occurring from local harnesses to always-on cloud harnesses.
These cloud harnesses can be accessed via Slack, mobile apps, and tunnel to local machines.
They operate in cloud sandboxes, enabling greater parallelism and resource utilization.

"this movement from a local harness to a cloud harness where the harness is always on."

The Harness to Claw Transition [07:00]

The transition to "claws" involves imbuing agents with initiative and learning.
Initiative means agents can proactively engage with external feeds and have a heartbeat for periodic actions.
Continual learning allows agents to auto-improve based on their execution traces.

"imbuing these agents, imbuing these harnesses with initiative and and learning"

Steinberger's Law: Every Harness Expands [09:49]

The speaker proposes "Steinberger's Law": every harness will expand until it becomes a claw.
This expansion is driven by user desire for more functionality, interaction channels (like Slack and text), and the psychological reward of getting useful outputs.
The desire for these agents is likened to a "dopamine casino."

"I believe every harness will expand until it becomes a claw."

The Future Shakeout [10:47]

A future shakeout is anticipated in the AI agent market, similar to the mobile app landscape of the 2010s.
Categories like ride-hailing or food delivery consolidated to one or two dominant players due to limited space in user cognition.
Agents will need to be either economically valuable or very frequent to remain pertinent in users' lives.

"after this phase where where we're sort of making everything more and more powerful um there will be a shakeout"

Implications for Builders [14:03]

Builders must stay updated with the rapid rate of change in the AI field.
Agents need to have the capabilities users require, as users may switch to more powerful alternatives.
Developers should anticipate future waves of consolidation and innovation.

"if you're building an agent, make sure that it has the capabilities that your users need"

Conclusion and Speaker Introduction [14:58]

The speaker, Sam, is the co-founder of Monsterra, a TypeScript agent framework, and author of "Principles of Building AI agents."
He encourages attendees to connect and discuss the topics further.

"it's great to see all of you. Thank you all for coming out."