Autonomy now mostly exists in execution …
Self Evolving AI Agents
The next step is to extend autonomy to self improvement…
Self Evolving AI Agents sound like something that is far off, but there are a number of elements of AI Agents which are already highly autonomous…
The one aspect is obviously the workflow / Agentic Workflow or planning the agent does.
There is an interesting study from OpenAI where they took training material and call data from a call centre and creating workflows via an AI Agent. Just by having the humans adapting their way of work to these agentic workflows yielded significant productivity gains.
The other aspect of self evolving AI Agents is an approach from NVIDIA where they created a framework for a continuous data feedback on which a Language Model is trained for improved Tool selection.
So it is a self-improving feedback loop for the AI Agent…
Another example of an AI Agent evolving is the new freeform function feature of GPT5.
The way I see this, is freeform function calling takes us closer to AI Agents that can write code to fulfil a task or a user request.
What aspects of an agent should evolve?
When should adaptation occur?
And how should that evolution be implemented in practice?
GPT-5 Freeform Function Calling Enabling AI Agents To Write Code to execute a task.
Considering the image below, you can have code tools/functions, flow tools, MCP integration, or Code Writing Tools. So if the AI Agent receives a request, it can actually write to code to execute the query.
I give a practical Python Notebook example here.
Another aspect that can be considered as dynamic and evolving is Memory…
Context is the working set of information actively used in the moment (at inference) to generate a response.
Memory is typically the stored data.
Memory informs context.
Back to the study…
The domain of the AI Agent should be kept in mind, companies like Anthropic, Xai, OpenAI and others build Agentic Solutions which is primarily General Domain.
For enterprises and SME’s there is a requirment for specific domain AI Agents. And hence SDKs, frameworks or scaffolding are required to build those solutions.
For self evolving AI Agents there are evaluation goals and metrics, like retention, adaptability, efficiency and more; which is shown below on the left.
But then there is also an evaluation paradigm…some assessments can be static, short or long horizon.
I have touched on the evolution of models, tools and memory (context)…the study breaks it down into what to evolve, when and where…
The schematic below under what context, models, tools, and architectures.
It also looks at the evaluation portion which is important.
The lastly, the image below shows the progression from LLMs to foundation Agents. And the loop to self-evolving Agents.
We currently find ourselves now in the “Execution via Tools & Planning” phase…and companies like NVIDIA is pushing the boundaries on Learning from Feedback & Experience.
Chief Evangelist @ Kore.ai | I’m passionate about exploring the intersection of AI and language. Language Models, AI Agents, Agentic Apps, Dev Frameworks & Data-Driven Tools shaping tomorrow.
