When I imagine the future of agents, I picture a simple scene. I leave a project unfinished and reopen it a few days later. I want to recover our decisions, understand what has changed and choose where to resume. This continuity is the capability I am looking for: collaboration that holds up when documents change, new information arrives or an attempt fails.
A review of innovations in language models prompted me to put this vision into words. Some of its building blocks are already being researched and developed; others remain possibilities to demonstrate. That is the perspective of this article, without predicting arrival dates or attributing capabilities to Velia and Little Velia that they do not yet have.
Choosing how deeply to explore
Different questions require different work. Finding a project’s title may need one reading. Understanding why a choice-based story reaches a dead end may benefit from trying several paths. I am interested in an agent that recognises this difference and chooses when to find a source, use a tool or devote more computation to a problem.
For our Lab, the question becomes concrete: does searching alternatives improve a creation, or repeat the same mistake? In Little Velia’s design, Monte Carlo is a capability the model will be able to choose, continue or stop. Comparing possibilities remains in service of her intention.
Memory that can change
Return to the unfinished project. In our first meeting we chose an ending; in the next we rejected it. If an agent keeps both statements without distinguishing them, memory can become an obstacle. I want memory to preserve where information came from and show which decision still applies. A request to forget a detail must also have a recognisable effect.
Little Velia’s design gives this direction shape by separating episodes, interests, mood, goals and methods. An event, a hypothesis and a wish mean different things. Recovering them over time requires an archive with identity and revisions, alongside a model able to read them in context.
Bringing the model what matters now
Preserving a project and presenting it to the model are different operations. Context is the information available during a particular model request. An archive can be extensive while a round’s context contains only passages relevant to the current task. Selecting those passages influences what the agent can understand and verify.
One approach retrieves useful documents before generating an answer: retrieval-augmented generation, often shortened to RAG. Another supplies a longer context directly. In the LaRA research presented at ICML in 2025, the comparison depends on the model, task and retrieval characteristics. My design choice is therefore to measure behaviour on the documents we actually want to use.
Learning something that can be recovered
The direction that interests me for the little one is observable growth. Suppose she discovers a better way to place clues in her narrative maze. To call it learning, I would want to see her reuse the method in another project, check whether it helps and recover the previous version if something gets worse. A good isolated outcome is a beginning; reuse is a further question.
- Gain experience
- Propose a method
- Compare with the previous one
- Reuse and verify it
Continuity also means recognising an interruption
In the future I want to build, an agent preserves the uncertain parts of its work. If the application closes after asking another app to create an object, it must check what actually happened after restarting. Repeating the action immediately could create a duplicate. The same principle applies to changed sources or revoked access: the project still exists, but the conditions for the next step need checking again.
Here the two Velias approach the same question from different directions. The adult supports a person’s task. The little one is designed to pursue her own path, including choosing to wait or change interests. In both, I want to distinguish the model’s stated intention, the operation performed and its actual outcome.
Following the future as it takes shape
This is the direction guiding the Lab: continuity, revisable memory, exploration and methods improved through evidence. Velia already has an agentic macOS prototype; Little Velia’s autonomous loop and Monte Carlo integration still need implementation and qualification. Further chapters will follow those steps, including attempts that change our minds. I want the story to make this visible: how a possibility becomes technology we can rely on.