Even when the developer draws upon application programming interfaces to forge connections to existing AI agents and services that provide essential functions, there’s always a huge amount of authentication and management that has to be applied at the infrastructure level.

Learning From Agentic Mistakes

Where the developer community may be falling short is in its ability to systematically track not just how and where agents forge their connection points, but also… crucially, to capture and codify what the agent has done through the course of its interconnection to another IT service and how successful that connection was in terms of the functionalities extracted from it.

“AI agents typically make the same mistakes repeatedly… which is a good thing in terms of learning,” explains Soham Ganatra, CEO of Composio, a newly established organization that works to provide a foundational skill infrastructure service for AI agents to learn and optimize their performance across enterprise applications. “When developers connect agents to applications like Salesforce (or a GitHub workflow, or a database), they all end up solving the same issues and providing the same context. We store this knowledge as ‘skills’ that can be reused by other developers. When an agent learns how to work with Salesforce for the first time, it shares that knowledge with every other agent interacting with Salesforce. This way, all agents get better over time.”

Now bringing its platform to market in its inaugural year of operation, Composio claims to be able to address a gap in the AI ecosystem by providing the "skill infrastructure" that allows autonomous agents to interact more intelligently with software tools and workflows. The company says that we can think of its services portfolio as an adaptive skill layer that improves its intuition with every interaction, that self-optimizes its own actions and becomes more useful with time.

Raw Intelligence Is, Well, Raw

Ganatra asserts that raw intelligence on its own (i.e. the power of large and small language models when fed through the most sophisticated AI data models and turbo-charged by the zippiest graphical or neural processing units known to humankind) is only part of the puzzle. He suggests that business impact happens when intelligence can learn from its interactions with the world, incorporating both the human beings and the machines that reside in it.

“AI agents will keep getting smarter on the backs of increasingly powerful models, but the magic lies in giving them a soul: elevating them from tool-calling robots to systems that understand your goals and your environment, growing with you as partners,” says Ganatra. “We’re building a self-optimizing skill layer, addressing the fundamental gap between increasingly intelligent LLMs and agents that can evolve from experience to develop nuanced, practical skills.”

He further states that skills are the fundamental building blocks behind an agent’s capabilities, ranging from basic tasks like email composition to complex operations like advertising management. An evolving skill layer improves the system's overall capabilities exponentially, allowing it to handle increasingly complex problems without extensive prompting.

The company says when building AI constructs, we need to think about the “scale of tools” i.e. for an agent to be reliable and skilled, it must access any software-as-a-service endpoint on demand. This warrants a repository with thousands of toolkits. In terms of skill adaptation, repeated workflows should transform from LLM-guided execution into fast, codified routines, optimizing speed and reliability. Skills should also dynamically switch between codified flows and LLM reasoning based on the context... this means that if a codified skill doesn’t exist for a task, it returns to LLM-guided execution.

Through the agents Composio across its customer base, the company says it is building the a network of managed skills that are personalized and learn from the environment that they operate in.

"You can spend hundreds of hours building LLM tools, tweaking prompts, and refining instructions, but you hit a wall," says Ganatra. "These models don’t get better at their jobs the way a human employee would. They can't build context, learn from mistakes, or develop the subtle understanding that makes human workers invaluable.”

Ganatra and team insist that they started building AI agent services “before the current hype” around two years ago. The business was founded on a mission to solve fundamental infrastructure problems in the AI space. The team tackled complex challenges including multi-agent coordination, authentication across enterprise systems and building scalability into infrastructure processes so that millions of requests could be handled daily. The organization now has over 100,000 developers using its platform, with tens of millions of requests every day.

Competitive Analysis, Agentic Automation

While Creatio will of course claim “uniqueness” and describe its agentic AI infrastructure technologies as one of a kind, we can find a range of services that reflect and resonate with its approach through the usual suspects in the IT marketplace… and within some of the start-ups vying for more voice in this arena

As well as Dynatrace with its Davis AI service and IBM with Watsonx Orchesrtate, the agentic infrastructure management space is well-populated with technologies from New Relic, Cisco, Splunk and even ServiceNow (as a firm that most of us recognize for its IT service management platform) with its technology that aims to empower individuals to become orchestrators of their own processes and provide agentic power to drive service discovery and incident resolution.

Among the big players, database vendor Oracle is in this sector with its Agent Studio, a technology for creating, extending, deploying and managing AI agents and agent teams across an enterprise. Arguably more narrowed to specific Oracle use cases, the company provides similar “codification of agents” with its agent template libraries, a service used to create AI agents with pre-built templates paired with natural language prompts, or draw upon a library of ready-made templates to support a variety of business scenarios. Also from the tech behemoths, Microsoft AutoGen is an open source programming framework for building AI agents and facilitating cooperation among multiple agents as they are engineered to solve workplace tasks. Microsoft says that AutoGen aims to provide a flexible framework for accelerating “development and research on agentic AI” with an emphasis on code quality, robustness, generality and scalability.

Lesser-known (but arguably no less tasty) fare is on offer in the space from Tonkean with its AI Front Door technology, which acts as an orchestration hub between workplace tasks and agents under enforced policy control with auditability. Again, there is a more narrow application point here i.e. Tonkean works specifically in the legal and procurement business. A little broader, Camunda is known for its work supporting the hybrid orchestration with agentic technologies. The company has been said to “meld” deterministic process models with non?deterministic AI?guided decision-making. MuleSoft, AWS, UiPath and others all operate in this space.

No More Back To School?

There’s something of an incongruous paradox going on in artificial intelligence if the drive to build agentic functions (which essentially exist to automate our lives with functional shortcuts and smart accelerators that remove grunt work) involves a core need to step on the brakes and go back to school for every new build. If AI is meant to encapsulate automation and make things quicker, then surely it should make use of encapsulated automation itself when attempting to operate. This is by no means the mission statement that Composio operates from, but it surely could be.

A core issue to address is the fact that language model interfaces have traditionally been built for humans to use, not for machine services to dovetail with. This means that AI in your favorite chat, email, search or creator application is fun to use, but it also means a hell of lot of work is going on below decks down in the engine room. Nobody wants to get covered in grease when they don’t need to, so let’s spare a thought for the AI agents themselves. Otherwise, bring some handwash.

The above is the detailed content of Composio Lays Down Score For AI Skills Infrastructure. For more information, please follow other related articles on the PHP Chinese website!