the question before the intervention
What would an agent need to know to occupy a role coherently? The actor’s goal, the causal model of the situation, current evidence, authority, commitments, constraints and available capabilities all matter. So does the unanswered question whose answer would change the action.
The context fabric research is about preserving and projecting that world. A search result is useful only when an agent can locate it relative to a decision. An apparently small question can expose the critical missing fact: what is true in this jurisdiction, what has the partner accepted, what evidence would change our belief, or what is the next action actually authorized to cause?
This line connects the public-data machine, which makes external facts citable, with the institutional kernel, which says which facts and actions a company accepts. It is a research program with working source systems, not a single finished product page.
the missing fact is often the whole job
An agent can retrieve a hundred documents and still ask the wrong question. Suppose an organization is considering whether to use an outside service. The archive contains specs, vendor emails and previous decisions. The action might depend on one unresolved fact: does the vendor’s current agreement permit the data flow this team plans to use? A longer summary will not solve that. The agent needs to identify the question, find the current authority for its answer, and change the plan if the answer is no.
This is what a context fabric should do. It carries the actor’s role, the situation’s causal shape, the current state, the authority boundary and the evidence that might change a decision. Retrieval supplies material. The fabric tells the actor what the material is for. Without that structure, a model can sound informed while acting on an obsolete promise or a fact true for a different jurisdiction.
a role is a view over the company
A sales actor, a finance actor and a support actor can look at the same customer episode and need different parts of it. Sales needs the promise made and the relationship history. Finance needs the authorized charge, refund state and settlement record. Support needs the customer’s current problem and the remedy it may offer. A useful role brief projects the shared world into those needs while keeping the underlying identities and events connected.
The projection should include the actor’s purpose, inputs, outputs, upstream and downstream dependencies, authority, commitments and the point at which it must escalate. That is richer than a persona paragraph. It tells an agent how its work can affect the company and where its power ends.
questions can be interventions
A consequential question is one whose answer changes the next move. “What was the most recent message?” may matter. “Has the customer accepted this remedy?” may matter more. “Which observation would make us abandon this plan?” can matter most of all. The point is to search for decision-changing information, not to ask clever questions for their own sake.
The public-data machine helps when the missing fact lives in an external record. The Institutional Kernel helps when the question is what the organization has accepted or authorized. Accountable history helps when several actors saw different stages of the same event. The context fabric joins those sources around a live role and a live choice.
The research challenge is to make that assembly reliable as roles, evidence and plans change. A useful test is to hand a new agent a real but bounded task. Can it state the one fact it still needs, the source that could answer it, and what it is allowed to do before the answer arrives? If it can, the fabric has carried more than text.
more context can make judgment worse
There is a practical limit to how much of a company an actor should see. Dump the whole archive into the prompt and old, vivid material can crowd out a small current fact. The fabric needs selection rules tied to the task: what affects this decision, what is current, which source has authority, and what uncertainty remains.
It also needs a way to mark contradiction. Suppose a contract says one thing, a recent email suggests another, and the operator remembers a verbal exception. A weak context system picks the most fluent story. A useful one presents the conflict and asks which source can legitimately change the agreement. The agent’s next move may be a question to the counterparty rather than an intervention in software.
The same discipline applies to memory. A past decision can be an example, a precedent, or a binding rule. Those are different roles. The context fabric should preserve the distinction so a helpful precedent does not quietly become policy. When an agent asks “what would change our next action?” it is really asking what authority and evidence the situation is missing.