Today’s models can reason across much of the world. But the agents built from them remain strangely unchanged by the worlds in which they actually work.
They encounter your codebase, your standards, your decisions, your corrections, and the exceptions no benchmark anticipated. Then the session ends. The next agent returns with the same general intelligence and too little of the specific intelligence the work produced.
We think this is temporary.
RSI has been looking in the wrong place.
Recursive self-improvement is usually described at the level of the model: a system that rewrites its own code, trains a more capable successor, and repeats. That is the most expansive form of RSI. It is not the only useful place recursion can begin.
An agent is more than a model. It is the complete system that shapes behavior: context, memory, instructions, skills, tools, workflows, evaluation, data, and the machinery that decides how those components should change.
When that deployed system learns from experience, the agent improves. When the machinery responsible for finding and evaluating those changes can itself improve, the loop becomes recursive.
The model is one part of the intelligence system. It is not the only part capable of learning.
work · constraints · outcomes
the operational definition of good
what changes, where, and why
context · memory · instructions · skills · tools · workflows
general capability
Qova works across the system around the model today. Model adaptation is a later consolidation layer.
Repetition is not recursion.
A system can run the same improvement recipe a thousand times without becoming better at improving. Memory is not recursion. Prompt optimization against a fixed evaluator is not recursion. Repeated reinforcement learning against a permanent reward is not recursion.
The threshold is crossed when the improvement mechanism becomes part of the modifiable system: when experience can change not only behavior, but how future behavior is judged, which evidence is trusted, where a change is encoded, and how subsequent improvements are proposed.
Behavior changes
adaptationEvaluation changes
adaptive learningThe improvement process changes
recursionThe definition of better cannot stay fixed.
Real environments are non-stationary. Teams change. Repositories change. Constraints change. Standards become more demanding. A learning system built around a permanent rubric eventually becomes excellent at optimizing a world that no longer exists.
The hard problem is not merely optimizing a reward. It is maintaining an operational definition of good that remains grounded as the environment moves—without allowing the learner to make its own success easier to claim.
This makes evaluation central. Outcomes, corrections, constraints, held knowledge, and the structure of real work must continually inform what survives.
Fast
Improve the deployed agent.
Experience changes context, memory, instructions, tools, workflows, and operational behavior.Slow
Improve how the agent improves.
Repeated evidence changes evaluation, reward logic, and the policy that selects future interventions.Personal is an ownership boundary.
Personal RSI does not mean small RSI. “Personal” identifies the world to which the intelligence is accountable and the boundary inside which its learning belongs.
That world can begin with one person. It can expand to a team, where useful patterns strengthen the group without indiscriminately pooling proprietary work. It can become organizational: an accumulating body of evaluations, datasets, and learning curves that remains owned by the organization that produced it.
Over time, sufficiently stable organizational learning can be consolidated into model weights. But recursion does not have to wait for that final layer.
What Qova Labs is researching
- Deployment-grounded learning
- How real work can produce durable improvement without turning every interaction into indiscriminate memory.
- Adaptive evaluation
- How the system can learn a changing operational definition of good while staying grounded in the environment it serves.
- Improvement allocation
- How a learner determines whether a lesson belongs in context, memory, instructions, skills, tools, workflow, evaluation, or eventually weights.
- Collective learning
- How individual learning curves can strengthen a team while preserving organizational data and access boundaries.
- Recursive learning policy
- How the machinery producing future improvements can become progressively more capable without losing external grounding.
The case for personal RSI is not that full recursive self-improvement has arrived.
It is that intelligence can begin improving recursively before the model learns to rebuild itself.
It can begin in deployment. It can begin with your agents. It can begin inside your world.
See QovaResearch context
Our framing draws on work in continual reinforcement learning, meta-reinforcement learning, adaptive reward models, meta-gradient optimization, and self-improving agent systems.