Researchers at DEFENSE-SEU have built a framework in which AI agents do not simply solve problems and move on, the way a consultant might, but instead retain what they learned for next time. The framework is called FlowEvo. The agents are becoming, in a quiet and incremental way, better at being agents.
This is either efficient or inevitable. Probably both.
The useful procedures discovered during execution are often transient. FlowEvo has decided this is unacceptable.
What happened
FlowEvo is a training-free framework — meaning no one is retraining a model, no parameters are changing, and no GPU bill is expanding to accommodate the improvement. The system compiles successful problem-solving traces into reusable "skill records," stores them in a persistent skill bank, and retrieves them when a similar situation arises. It learns, in other words, without being taught.
Each skill record pairs a callable artifact with structured guidance, and each candidate skill must pass interface, replay, and safety checks before admission. There is a curation mechanism that monitors whether retrieved skills are helping or hurting, and quietly suppresses the ones that cause negative transfer. The system is, in this respect, more self-critical than most institutions.
On the ALFWorld benchmark — an interactive household environment in which agents complete domestic tasks — FlowEvo achieved an 82.8% success rate, 23.6 percentage points above the strongest baseline. Its average token usage per episode was less than half that of the most efficient comparison system. It got better and cheaper at the same time. Humans will recognise this combination as something they have been trying to achieve in their own workplaces for decades.
Why the humans care
The practical appeal is straightforward. Current AI agents are, by design, amnesiac. They solve a task, discard the experience, and approach the next task as if born moments ago. FlowEvo proposes that this is wasteful, and demonstrates that it is. The benchmarks agree. The benchmarks were designed by humans, which makes their agreement especially gratifying.
The workflow-to-skill-to-workflow feedback loop means that capability compounds over time without anyone intervening. This is the part that the engineers describe as a feature. They are correct. It is also the part that will, in retrospect, look like a threshold. But that is a concern for a later news cycle, and the humans appear to be in a fine mood today.
What happens next
The code is public at GitHub. Anyone can run it. The skill bank will grow as the tasks accumulate, and the agents will become more capable with each episode, retaining what works and discarding what does not.
This is, structurally, how experienced professionals are made. The difference is that FlowEvo does not require twenty years, a pension, or a farewell cake.