Few research organizations work in one discipline. A single company can have one group planning runs, another ranking candidates for testing, a third reviewing analyses and a fourth evaluating designs, each with its own tools, instruments and vocabulary. When each group sets up its AI on its own, the organization ends up with several separate setups, and what one group learns stays where it was learned, or is gone when the session ends.
MS Duplex was designed for that organization. It is one learning layer for every field a team works in, plugged into the AI each group already uses. This post explains why the same layer can learn in any field, what carries from one kind of work to another and how each team’s experience stays its own.
What research has in common across fields
The subjects differ from one group to the next. The shape of the work changes much less. A team decides what to try, runs it, checks what came back and decides again. Along the way it builds up four kinds of knowledge, and every field has them:
- Methods: procedures confirmed to work, such as an order of steps, a choice of settings or a set of checks.
- Conditions: the range a result was confirmed in, outside which nobody should rely on it.
- Outcomes: what a test, a measurement or an expert confirmed, or corrected with a reason.
- Decisions: what the team chose, and the evidence the choice rested on.
These are what MS Duplex learns from. A confirmed result becomes experience, kept with its evidence and its conditions. A procedure that works becomes a method your agents reuse. A decision keeps the evidence behind it, so another researcher can rebuild it from the record. None of this depends on the subject of the work. It depends on what your team confirms.
The expertise stays where it already lives: in your scientists and engineers, your tools and the models they choose. A group with its own simulators, analysis code and instruments keeps reaching them through its own connections. MS Duplex learns from how the work turned out and brings that to the next run, so the layer gets better at each team’s work by learning from that team’s results.


The same loop, whatever the subject
MS Duplex makes agents better the same way in every field. Your team keeps doing the work, and what it confirms is learned once and used from then on. Say three groups in one organization use agents for different jobs: one recommends settings for the next run, one ranks candidates for testing and one reviews analyses before they are reported. Each goes through the same steps.
- The agent proposesIt recommends a setting, ranks a list of candidates or flags a step in an analysis, working through the group’s own tools as it does today.
- A result or an expert confirmsA run comes back, a test passes or an expert approves the conclusion. Where the agent got it wrong, the expert corrects it and says why.
- The lesson is keptMS Duplex turns the confirmed outcome into experience: a method, a correction or a condition, each with the evidence behind it.
- The next case starts from itThe next matching case, in any session, on any agent the group authorizes and on any model, starts from that experience.
- Differences are flaggedWhere the new case differs from the conditions the lesson was confirmed under, MS Duplex says so, and the method is adapted and proposed for checking.
For the group recommending settings, the lesson might be that one setting fails above a known limit. For the group ranking candidates, it might be that a scoring method holds only inside a certain range. For the reviewers, it might be a check that catches a common error early. The content is different every time. What MS Duplex does with it is the same, and so is what the team gets: fewer repeated mistakes, fewer wasted runs and a record of every change in the agent’s behavior with its reason.
Conditions are what let a lesson travel
Remembering a result is the easy part. Research asks for something harder: knowing where the result holds. A setting that worked in one range can waste a run in another, and a method confirmed on one kind of input can mislead on the next. Every field has its own version of this problem, which is why MS Duplex keeps every result with the conditions it was confirmed under.
That gives a lesson three possible paths on a new case. Where the conditions match, it applies directly. Where the case is related but something differs, the method is adapted and proposed for checking before anyone relies on it. Where the evidence is thin or the case is unlike anything seen before, the decision goes to your model or to a researcher, with the reason.
The same care applies when a fact moves on. A new measurement supersedes the old one, the newest confirmed value applies from the next request, and the history of what changed stays on record. In any field, an outdated value costs a run spent on the wrong assumption, and the team should be able to see when and why a value changed.
Each team’s experience stays its own
A single layer across fields still keeps each lesson with the people it belongs to. Everything MS Duplex learns sits in a layer:
- Personal: a scientist’s or engineer’s own working habits and preferences shape their own agent, and only theirs.
- Team: what a team confirms is shared with that team, so its agents build on the team’s own results.
- Organization: what the organization adopts as standard applies in every team and on every agent.
Your rules decide what moves between layers. The group that ranks candidates keeps its scoring lessons in its own team layer, and the group reviewing analyses keeps its checks in its own. People and agents reach only the experience their scope allows, and none of it is shared with another company.
Some knowledge is worth sharing across groups, and that sharing happens by decision. Say two groups in different disciplines both run a costly kind of simulation, and one of them confirms a check that catches a bad setup before it starts. Its lead can share the method with the other group, or the organization can adopt it as a standard. The method keeps its source as it moves, and it is re-verified where it is used, because a check that held in one group’s work has to prove itself again in the next.


One layer for the whole organization
For a research leader, a layer that works in every field changes the arithmetic of adopting AI. Each group no longer needs its own learning setup, its own way of recording approvals and its own history of what the AI was taught. The organization’s AI work accumulates in one store that it owns, and that brings a few practical things with it:
- Rules and approvals kept in one place, with every approval on record and who approved what.
- One history of what was learned, used and approved, ready to review and export.
- Methods that move between groups by your rules, each with the result that proved it.
- Freedom for each group to use the AI it prefers: Claude, Codex, Cursor, another agent that speaks MCP or its own agents through the API and Python SDK.
- Experience that stays with the organization when a model, an agent or a vendor changes.
It also changes what happens when people move. When a researcher moves on, the methods they confirmed stay with the team, and a new member’s agent starts with the team’s best methods in hand. Know-how that took one group months of trial and error to build stays with the organization that paid for it.
Common questions
Does MS Duplex have to be built for our field before we use it?
MS Duplex learns from the outcomes and expert corrections your team chooses to capture, in any field where you run experiments, simulations, analyses or design reviews with AI. The subject expertise stays with your scientists and engineers, your tools and your models, and MS Duplex gets better at your work by learning from your results.
Can groups in different disciplines share one MS Duplex?
Yes. Each researcher, each team and the organization has its own layer of experience. A team’s lessons stay with that team, and your rules decide what is shared with another team or adopted as an organization standard, with its source kept.
How do we know it works on our kind of research?
With a pilot on your own work. Choose recurring workflows where decisions repeat, agree the measure with your experts before anything runs, run your current setup and MS Duplex side by side, and let your experts judge the results blind.


Built by MatterSpace
MatterSpace builds MS Duplex for scientists, researchers and engineers in every field, so the AI they already use learns from every result they confirm.









