External memory can help an agent do better work without retraining its model. But more saved experience does not, by itself, make it a better learner or justify giving it more decisions to make. In a company, an agentic workflow can use AI to gather evidence, explore possibilities, and handle tasks within clear boundaries. The goal should be to expand what people can understand and decide, not assume they want to leave every decision to an agent.
What changes when the model stays fixed?
When a model is not being retrained, its weights—the learned numbers that shape how it processes information—stay fixed.
The harness, the software around it, can save experience, revise Skills, and choose what to bring into the next task. Better guidance can improve the system’s work. But saving that guidance does not itself train the model that interprets it.
That is where Part 3 left us: fixing a lesson and improving how the next lesson is formed are different changes. Experience can help a person notice new distinctions or reconsider how they draw conclusions. A growing folder of instructions does not establish the same development in an agent.
The system may improve through changes outside the model too. We still need to show which judgments improve through experience, rather than assume more records will produce that growth.
What would a continually learning agent look like?
Richard Sutton’s OaK architecture explores a system whose components keep learning from experience. It learns to predict what actions will lead to, then uses those predictions to plan.
It also proposes learning how strongly individual weights should update. In plain terms, experience would help shape both what the system learns and how much it changes after new evidence. That reaches toward the question from Part 3: what updates the way it learns?
This goes beyond writing another lesson for a fixed model to read. It is still a research direction, not proof that we can build an assistant that reliably develops through years of open-ended work.
Why should more autonomy be the goal?
Even if we build agents that keep learning, we still have to decide what we want to delegate. Being able to make a decision does not automatically mean being authorized to make it.
That raises another question: why assume the best agent is one that leaves us with nothing to do?
Wanting less busywork does not mean wanting less say.
People can want help gathering information and still want to understand the choice. They can want several tasks handled automatically and still care about setting the direction.
AI can expand that participation. It can gather evidence we would struggle to find, suggest explanations we missed, and help test our assumptions. It can make the trade-offs easier to see before we decide.
That is a valuable direction for agent development. Success need not be measured by how completely the person disappears from the process.
Giving a system more independence also transfers decisions to it. Removing a review step does not make it better at the judgments that the reviewer supplied. More memory alone does not justify that transfer.
In a company, start with the process
When a business process already has clear rules and decision rights, the design should start there.
For invoice review, an LLM might extract fields from a messy document and investigate unclear details. Software can check totals and explicit policy rules. A person can decide exceptions that require business authority.
An agentic workflow can give the model room to choose tools and handle uncertain steps within those boundaries. Memory helps preserve relevant context. The process makes clear what the model can decide and when it must pass the work on.
Building a society of agents does not answer those questions. Neither does giving them access to every company document. More information does not determine who should have authority.
Autonomy should serve the process. It should not become the purpose of the system.
What are we expecting from the tool?
This series began with a question: how does yesterday’s experience change today’s decision?
Saving it, recalling it, and turning it into a Skill all help. These achievements alone do not show that we have built a system that keeps becoming a better learner, let alone establish a path to AGI.
And even if that capability becomes possible, we would still need to decide what we want to delegate.
Giving an agent a memory does not make it a human-like learner. And making AI more capable does not mean people should have less say.
The point of the tool is to expand what we can understand, explore, and decide.
Resources
The article originally appeared on LinkedIn.