Do Large Action Models (LAMs) Really Exist?
Artificial intelligence has long been associated with language-based interactions, thanks to Large Language Models (LLMs) like ChatGPT and Bard that excel at generating human-like responses. However, as AI advances, a new term has started making waves—Large Action Models (LAMs), which are supposed to move AI beyond just reasoning and text generation into autonomous execution. But do they actually exist, or is this just another industry buzzword?
Unlike LLMs, which specialize in understanding and generating language, LAMs are described as models that can plan and execute real-world actions autonomously. The claim is that while an LLM might advise you on how to book an appointment, a LAM would actually book it for you by interacting with the necessary systems, APIs, and workflows.
The problem? There are no clear, publicly available LAM architectures today—only repurposed AI agents and automation tools rebranded as “LAMs.” A true LAM would require end-to-end autonomy (no human oversight) and generalized action learning (adapting to novel tasks through trial and error, not just executing scripted API calls). What’s being marketed as a “LAM” is often just a well-integrated AI agent, a fine-tuned LLM for structured actions, or an automation tool wrapped in AI branding. This means that while the idea of LAMs is appealing, the reality is that they don’t yet exist as a distinct AI category.
What’s Being Marketed as a LAM Today?
Several companies are pushing products labeled as LAMs, but on closer inspection, most of these fall into one of three categories:
Fine-tuned LLMs with tool-use capabilities
(e.g. models that generate API calls and structured actions, like Gorilla)
AI agents with execution capabilities
(e.g., AutoGPT, BabyAGI, Rabbit R1, Adept ACT-1—none of which are truly LAMs, but rather agentic systems with execution workflows)
Enterprise automation solutions
that claim to use “LAMs” but are actually RPA (Robotic Process Automation) combined with AI
Some of the most commonly referenced “LAMs” include:
Rabbit R1
A consumer AI assistant marketed as using LAMs, though it remains closer to an AI agent than a full-fledged LAM. Early reviews, such as The Verge’s critique, highlight its unreliability in real-world tasks like food delivery.
Adept ACT-1
This AI model interacts with software tools and APIs, demonstrating some LAM-like behavior, but it’s functionally an advanced AI agent—not a new category of AI.
Salesforce’s xLAM
A research-stage framework for multi-agent execution, still under development and not yet a deployable model.
Microsoft’s Research
Work on architectures like TaskMatrix.AI explores LAM-like ideas but remains academic.
While these systems show aspects of LAM-like behavior, they are still built on existing AI technologies rather than introducing something fundamentally new.
LAMs vs AI Agents: Are they actually different?
One of the biggest problems with the LAM narrative is that AI agents already do what LAMs claim to offer. AI agents are designed to:
The supposed difference is that LAMs are meant to be “autonomous execution engines”, but the reality is that most current AI automation systems already achieve this through well-orchestrated AI agents.
This comparison highlights that most of what LAMs claim to offer is already achievable with well-structured AI agentic workflows, making the term largely redundant.
Rather than being a new category of AI, LAMs seem more like an evolution of AI agents, integrating execution more tightly into reasoning workflows. For instance, a true LAM would not just chain predefined actions but learn to execute novel tasks through trial and error, much like humans adapt to unfamiliar software.
The Reality: What Needs to Happen for LAMs to Be Real?
For LAMs to become a distinct and meaningful category, they would need to:
Break free from LLM dependencies
A true LAM should not just be an LLM that generates API calls. It should have a structured execution model beyond text-based interactions.
Learn from real-world actions, not just text
Current AI models are trained on text corpora, but an actual LAM would need datasets of multi-step decision-making and action execution (e.g., navigating a hospital’s EHR system without pre-programmed APIs).
Introduce true autonomy
Instead of requiring agentic orchestration, a LAM would have to independently plan and execute complex workflows based on its own reasoning.
Deliver generalizable execution
Unlike today’s agent-based AI, LAMs would need to adapt workflows across domains (e.g., a single model managing both supply chains and customer service).
Until these capabilities are clearly demonstrated, LAMs remain more of an aspirational concept than a functioning reality.
The Alternative: AI Agents & Execution-Optimized Architectures
Instead of waiting for LAMs to materialize, businesses today can achieve the same level of automation by leveraging AI agents with structured execution layers.
For instance, AI-driven automation solutions—like Superbo’s microassistants running on the Superbo GenAI Fabric framework—already offer:
Rather than relying on a vague new category like LAMs, AI-driven automation can already deliver practical, scalable execution capabilities today
Conclusion: LAMs – Future or Just Buzz?
The idea of Large Action Models is intriguing, but today’s reality doesn’t match the hype. Most of what is being marketed as a LAM is just an extension of existing AI agentic workflows—fine-tuned LLMs, automation scripts, and API orchestration layers wrapped in a fancy new label.
At this point, LAMs are not a fundamentally new AI category. They may become more distinct in the future, but for now, businesses should focus on leveraging AI agents, execution-based architectures, and structured automation workflows to achieve true AI-driven efficiency.
Superbo’s microassistants allow for the use of execution-focused AI as part of their microassistant design and architecture, without relying on overhyped, non-existent AI categories. Instead of waiting for LAMs to evolve into something real, businesses can already achieve true AI autonomy today—without the marketing fluff.
If and when LAMs become a truly distinct AI category, the industry will know it—not because of marketing, but because of tangible technological breakthroughs. Until then, businesses should focus on what actually works today.
