In this article
An agent becomes useful to a company when it understands enough about the work to do it well.
It needs to know what the numbers mean, which sources to trust and how decisions get made. It needs the methods people have developed through experience, including the exceptions that rarely make it into a handbook.
This is what we mean by context. It shapes both what an agent knows and how it works.
At Sento, we believe getting that context right is how we build agents with skills specific to a company. It is also how we make those skills available beyond the person who first figured them out.
From individual progress to shared skills
Much of AI adoption still happens alone. Someone finds a useful task, gathers the information and works through a prompt until the result is good. Their colleague goes through the same process the next day.
Both people benefit. But the company has no dependable way to build on what they learned.
We want to change that.
Stripe’s walkthrough of Kai shows what this can look like. Kai combines knowledge about the employee with projects, tools and reusable skills. Project owners can configure the models and permissions appropriate for the work.
One moment in the demo captures the idea particularly well. A user builds a dashboard through conversation, refines it, then turns the session into a skill. Kai can later retrieve that skill and repeat the workflow.
The company has gained something it can use again.
That is close to our ambition with Sento. We want teams to build agents that become good at their particular work, and give them a shared place to develop and manage that capability.
Context gives agents their specialisation
Sento’s Foundation brings together company knowledge, definitions, metrics, instructions and skills. People can maintain that context and make it available across their connected AI tools. The broader product brings the skills, routines and agents built on that foundation into one place to run and improve them. Learn more about Sento.
The context is what connects those pieces.
Consider an agent helping with customer renewals. Knowing a contract’s end date is useful. Doing the job well also requires knowing how the company judges renewal risk, what was promised to the customer and which next steps need approval.
A skilled account manager brings those things together. Our ambition is to make more of that expertise available to an agent, in a form the team can inspect and improve.
The same applies to other work. An agent preparing a product update needs to understand the audience and what actually shipped. An agent investigating a metric needs to understand its definition and the limits of the data behind it.
These agents may use the same model. Their skills come partly from the different context, methods and tools available to them. Good context gives them a basis for making choices that fit the company. Testing their work shows where that basis still falls short.
The people who know the work shape the agents
This makes the people who know the work central to building useful agents.
The support lead knows which issues need special care. The finance team knows what belongs in a revenue figure. The account owner knows why an ordinary process would be wrong for a particular customer.
They need a practical way to contribute that knowledge without becoming responsible for an entire AI system. We have written before about how much company knowledge remains unwritten. Capturing it means making room for meaning, exceptions and judgment as well as facts.
Shared context also needs ownership. When something changes, the team should know where to update it and who is responsible. The useful correction should be available to the next agent that needs it, rather than remaining in one person’s conversation.
Agents can contribute to the company brain
Agents can collect new information and write results that other people or agents will use. A scheduled task might record a product release. Another agent might use that record to prepare a customer brief.
For that to work, the shared context needs clear rules. In Sento, writes are checked for permission, structure and, where relevant, the version being updated. Activity is logged so we can see what changed and who changed it. We describe this in more detail in our post about letting agents write to the company brain.
Those controls make it possible to involve more contributors while keeping responsibility clear.
This is what multiplayer means to us. People and agents can work on different tasks while contributing to a shared understanding of the company. Their work can become useful to someone they never directly collaborated with.
A better starting point for AI adoption
Shared context changes the starting point for adoption. A new teammate can use an established skill with context already maintained by the people who understand the work. They still need to exercise judgment, but they have less to assemble before they can get value.
There is no need for every agent to know everything. A company may have a skill that matters to only three people. It can be highly valuable in that setting. The goal is to give each agent the context and capabilities its work calls for.
Stripe’s Kai is a useful example of how far this thinking can go. For us, it reinforces why context sits at the centre of what we are building.
We want a team’s experience to become something its agents can work with. As that experience grows, the team should be able to develop more capable agents without repeatedly explaining the company from the beginning.
A person finds a better way to do something. The team makes it available as a skill. The next agent starts there.
