MS DuplexSelf-improvementShow allShow less
Self-improvement.
Your agents get better with every confirmed outcome.
Most AI agents behave on their hundredth task as they did on their first. MS Duplex makes the agents your team already uses improve with every round of work.
Learn and improve · one of the fourteen capabilities of MS Duplex


Your team gets
- Fewer repeated mistakes and wasted runs
- Agents that improve on their own between reviews
- A record of every change in behavior and its reason
What self-improvement means for your team.
Self-improvement means your agents change how they work because of what your work has confirmed. An approach that succeeded moves to the front of the list. A step that failed under certain conditions stops being proposed under those conditions. Each round of work makes the next one start further on.
The measure is your own outcomes. MS Duplex judges each change in behavior by the results your tests, measurements and experts confirm, and ties it to the outcomes that justified it. Because that evidence keeps arriving as the work goes on, your agents improve on their own between reviews.
Every improvement is on record. For each change, your team can see what the agent now does differently, since when, and which outcomes justified it. When a lead asks why the agent behaves better today than it did last month, the answer is there, with its sources.
Why research needs agents that improve with use.
Research work repeats. The same kinds of decisions come back week after week, and an agent that does not learn from them makes the same mistakes at the same cost.
Much of a research team’s AI work is recurring: ranking candidates, choosing settings, planning the next batch of runs, checking an analysis. Each round produces evidence about what worked and what did not. An agent that improves with use turns that evidence into better behavior on the very next round.
Without self-improvement, the improvement has to come from people. Experts review the same kinds of output and correct the same kinds of error, prompts and notes grow longer, and the agent stays as good as the last time someone edited its instructions. The team’s experience grows while the agent’s stays where it was.
Recurring work adds up
A small improvement on a decision your team makes every day is worth more than a large one on a decision made once.
Wasted runs are costly
Every failed approach an agent proposes again spends compute, materials or instrument time a second time.
Reviews have limits
Experts cannot check every step, so the agent has to get better between their reviews as well as during them.
Leaders need to see progress
An improvement your team cannot trace to evidence is hard to trust and harder to defend.


How self-improvement works in your team’s day.
Your agents keep doing the work. MS Duplex takes the outcomes your team chooses to capture and turns them into better behavior for the next round.
- 1
The agent proposes
It recommends an approach, a setting or a next step, drawing on what your team has already confirmed.
- 2
The outcome comes back
A test, a measurement or an expert shows whether the proposal worked, and under which conditions.
- 3
Behavior changes
What worked becomes one of the agent’s first choices. What failed under those conditions stops being proposed there.
- 4
The change goes on record
Each improvement is kept with the outcomes that justified it, so your team can review it, correct it or remove it.
An example, in any field where an agent plans runs.
A team’s agent plans batches of runs toward a target. Over several rounds of work, its plans get better, and the record shows why.
- 1
Round one
The agent proposes ten runs. Three reach the target, and four fail for the same reason: a setting outside the range where the method holds. The team confirms both results.
- 2
Round two
The agent starts from the approach behind the three successes and keeps the failing setting out of its plan under those conditions. Fewer runs go to what the team already knows does not work.
- 3
Round three
A run with a new combination succeeds where the old first choice fell short. The new approach takes its place at the front, tied to the result that justified it.
- 4
The review
At the next review, the lead opens the record. Each change in the agent’s plans is listed with the outcomes behind it, and one that rested on a single doubtful result is removed on its own, while the rest stays.
Self-improvement draws on the other capabilities.
Better behavior depends on good evidence, clear reasons and a record your team can trust.
Continual learning
Every confirmed result is learned from the next request on, which gives self-improvement fresh evidence with every round.
Learning cause and effect
Knowing which action led to which outcome, under which conditions, lets your agents improve in the right place and leave the rest alone.
Decisions that know their limits
As approaches prove themselves, more routine decisions are settled directly, with fewer model calls on recurring work.
Provenance, review and unlearning
Every change in behavior keeps its source, so any one can be reviewed, corrected or removed without touching the rest.
What your team gets.
Agents that get better with every round of work, with every improvement visible and tied to its evidence.
- more accurate than a model fine-tuned for the task, on recurring work, with no training run, in MatterSpace’s controlled tests
- +16 pts
Fewer repeated mistakes
A step that failed under known conditions stops being proposed under those conditions.
Fewer wasted runs
Compute, materials and instrument time go to approaches with confirmed evidence behind them.
Improvement between reviews
Agents get better on their own as outcomes come in, without waiting for someone to rewrite their instructions.
A record of every change
Each change in behavior is listed with what changed, when, and the outcomes that justified it.
Measured on your work
Improvement is judged against the outcomes your own tests, measurements and experts confirm.
No training runs
Your agents improve beside the model, with its weights left as they are, so there is no dataset to prepare or training cycle to wait for.
Common questions.
Make the next run benefit from the last.
Bring us the research your AI should get better at. We will plan a pilot on your own workflows, with your experts as the judges.

