X Lab is JoinX's research centre. The new world is the one we build for agents: a permanent layer where context, rules and memory live, and where the intelligence stays with the company even when the agents change.
We explore the frontier of AI, meaning the models, agents, protocols and tools that come out every week, and among so many new releases we pick those that could really change how companies work over the next twelve months. We put them to the test before they become commonplace, with a clear view of what really matters in organisations.
A frontier technology is not yet ready for a company: it lacks access rules, predictable costs, traceability and integration with the systems already in place. In X Lab we do this work, so what the market will bring to maturity in a year our clients use tomorrow, with the same guarantees as everything else.
TodayIn twelve months
The market · Ready for the company in a year
With X Lab · Ready for the company tomorrow
Illustrative diagram
AprogrammableworldforAI,infourdimensions
We treat context as a world that is programmed first and executed afterwards, and at every moment its state decides what each agent sees and what it can do.
Time
What is remembered from one step to the next, for how long, and what is forgotten
State
The condition a case, a system or an agent is in at that moment
Space
The place where the action happens: a system such as the CRM, or a physical place
Knowledge
The part of the company's knowledge that this agent can use, in that context
The harness lives in the world, and every agent inherits it
A language model does not remember: at each call it receives a text, responds and forgets. To make it work in a process, a harness is built around it, meaning instructions, tools, rules, access and memory, and today it is built inside every agent, which carries its own copy.
We move it into the world. The company's world becomes a permanent layer, described and governed, that all agents share, and each agent visits it reading only the part it is entitled to. So a rule is updated once, in the world, and applies at once to every agent.
Whatweworkon
These are the themes on which X Lab concentrates its testing and development. Each one comes from a problem we meet in projects.
01
The world as a model
People, agents, resources, contexts
We describe the company as a world of actors, resources and contexts, in which every link between an actor and a resource is also a permission.
02
Governed memory
Remembering the why
The scattered traces of a case are consolidated into structured profiles, each with its reason, and the memory forgets what is no longer needed.
03
From skill to process
Reasoning paid for once
A validated procedure stops being an instruction to reinterpret and becomes an executable process: the fixed steps consume no tokens, the judgement stays with the model.
04
Reasoning only when needed
The simplest level that is enough
Each request first looks for a process already known, then a composition of steps, and only at the end the model's free reasoning, the most expensive option.
05
Open standards
MCP, A2A, contracts
We connect agents and systems through open protocols, so even an agent the company already has enters the world without being rewritten.
06
Replaceable models
Sovereignty over data and knowledge
The same world works with different models and in different environments, from the company's cloud to an isolated environment, and it remains the property of those who built it.
Aworldthatgovernstheagents
Three pieces of X Lab work in which the result improves because the world the agent works in changes.
01 · From state to experience
Agents that learn by watching how the work is done
The world keeps the state of things and also the way they were done. Agents observe the runs, by people and by other agents, and recognise the paths that reach the result in fewer steps.
The first time
The agent reasons step by step, the most expensive way, and every attempt stays in the world as experience.
With use
The path that works is reinforced at every repetition, and a new goal that resembles one already seen starts from there.
Once validated
The path becomes a procedure that runs almost without a model. Reuse proposes, the rules of the context decide.
Illustrative diagram
02 · The harness in the world
No super agent, fewer tokens at every call
When rules, access and memory sit inside the agent, you end up building a single one that knows everything, and reloads everything at every call. With the harness in the world, small, specialised agents are enough, and the model comes into play only when needed.
A super agent
A single agent with all the context: at every call the model receives the entire archive, including what is not needed.
Agents in the world
The world passes each agent only the part of the context it is entitled to at that moment.
The model last
What is already decided runs without a model, and free reasoning is kept for new cases, the smallest part of the work.
Model
Tokens per call
Illustrative diagram
03 · Company Brain
A memory that consolidates and knows how to forget
The Company Brain keeps rules, definitions and decisions with the reasons behind them. A memory that accumulates everything soon becomes unusable, so ours stores, consolidates and forgets according to written rules.
Stores
Every conversation, case or decision leaves a specific data point in the world, with its provenance.
Consolidates
Scattered data is gathered into structured profiles, and what is used often stays close to hand.
Forgets
What is no longer needed fades out according to rules decided by those who govern the memory, and the model cannot change them.
Illustrative diagram
Whatisreadyisdecidedbythetests
Every technology that enters X Lab goes through a test bench, with thresholds set before the test. We propose to a client only what has met them, and the people who run projects already know what is worth using and what is not.
Many new releases do not deliver what they promise, and for us a negative outcome is as much a result as a positive one: it spares a client months spent on the wrong road.
X Lab · Test benchUnder testTests completed
Reading scanned documentsDocumentsPassedOur test thresholdHolds up on contracts and handwritten forms
Agent operating in the browserAgentsNot yetStalls on multi-step forms
Next-generation language modelModelsPassedHolds up on Italian-language texts, with sources cited
For project teams
Worth using
Reading scanned documentsNext-generation language model
Not yet
Agent operating in the browser
Illustrative example
Whatchangeswhentheharnesslivesintheworld
Anyone who puts a lot of AI into processes soon meets the same problems, and they depend on where rules, context and memory live.
If everything sits in the agentIf it sits in the world
AccessIf everything sits in the agentEach agent has its own permissions, and nobody oversees the wholeIf it sits in the worldAccess rules are part of the world, and the very structure of the system shows each agent only its own part
CostsIf everything sits in the agentAt every run the agent reloads all the context, and tokens grow with coverageIf it sits in the worldThe world projects only the live part of the context, so cost follows the work done even as the world grows
UpdatesIf everything sits in the agentWhen the company changes, each agent has to be updated by handIf it sits in the worldThe world is updated once and all agents see the new state straight away
ReuseIf everything sits in the agentEach new agent starts from scratchIf it sits in the worldEach new agent is born into a world already described and inherits rules, contracts and memory
TraceabilityIf everything sits in the agentThe more agents there are, the less is known about what each one didIf it sits in the worldEvery action carries its own why, written before acting
OwnershipIf everything sits in the agentThe context lives in the platform of whoever supplies the agent, and leaves with itIf it sits in the worldThe world belongs to the company, in its systems, and stays when agents and models change
Reasoningpaidforoncestayswiththecompany
When a procedure moves from free reasoning to an executable process, the company pays for it once and reuses it at every run, and what one project describes in the world also serves the next. This is Compounding Intelligence applied to agents: the world grows with every process, and with it what the agents can do.
No. It is our in-house research centre: what its tests confirm goes into the modules and reaches clients through the solutions, starting with Deployment.
What does harness mean?
It is the set of instructions, tools, rules, access and memory that keeps an agent at work inside a process. It is usually built inside each agent, whereas we build it once, in the world, and every agent inherits it.
How is it different from an agent framework?
Frameworks are there to make agents work, and we use them too. Our work sits one level above: the world that agents answer to, with its context, its access rules, its memory and its traces.
Is the world you talk about a world model?
No. Today world model refers to models that generate simulated environments. Our world is the layer in which the company describes people, agents, systems, rules and memory, and in which the agents work.
Can we bring a case of our own to X Lab?
Yes, starting with a conversation. It can be an agent you already have, a process to rethink or a technology to evaluate: we tell you what we have already tested and what would need testing.
Bringoneofyouragentsintoagovernedworld
We can start from an agent you already have, from a process or from a question about the frontier: we tell you what we have tested and what holds up.