Implementing context management: connect processes, skills and knowledge

Begin with the business procedure from the incoming request to the handled service case. In our example, a customer reports that her washing machine will not start. Customer Service records the available details. An agent is to perform the activity “Prepare the service handoff”; a person then reviews the draft and passes it to technical service. That team handles the case and brings it to a close.

In the process model, assign activities to their performing roles, for example in separate lanes for Customer Service, the agent and technical service. Define the input and expected result for the selected agent activity: the details received so far should become a case note that identifies missing information and names the knowledge used. The simplified diagram below shows this activity in the process and its connection to the external implementation.

For this first step, the agent decides neither on a repair nor on a service appointment. It returns a draft for review. An unresolved model identifier remains visible as a follow-up question; additional details can subsequently be used to prepare the note again. Roles in the process show who performs the work. Technical permissions and the mechanism that invokes the agent need to be configured separately during implementation.

The working session needs the perspectives of intake, subsequent case handling and the person responsible for the instruction. Whoever implements the agent activity also establishes how it is invoked and which connections it needs. Use a suitable anonymised case or an explicitly fictional one. Our example develops the distinction between context, AI assistance and business applications from my article “Wenn KI mitarbeitet”.

The starting note captures the request but leaves unclear what information the next stage needs. Work through the case with the person who will take it on after the handoff. Whenever they would need to ask a question, record what they need the answer for. This defines the requirements for the activity’s result before you describe the agent’s internal steps.

The model identification is initially missing, which prevents the appropriate service instruction from being selected reliably. “Will not start” also needs a more precise description before technical service can interpret it. Service determines which questions intake should ask from its approved instruction. The result is a better description of the case; a technical cause has not yet been established.

Now associate the relevant business knowledge with “Prepare the service handoff”. In the proposed design, the applicable service instruction and handoff rule are linked to this activity in PULSE. Employees and a suitably connected AI application can then identify the sources intended for this point in the process. The particular case, such as the model range involved, further determines the selection.

The relationship to the activity describes how the knowledge is used. It should be more precise than a loose collection of every document about service. Maintenance instructions can be technically correct yet unsuitable for fault intake by Customer Service. The people responsible therefore need to establish which activity, audience and model range the linked instructions apply to.

The customer report remains part of the individual case. It is supplied to the agent for this assignment or retrieved through a connection configured for that purpose. PULSE does not need to hold every case record. The maintained business context and current case details come together when the particular activity is performed.

If two sources prescribe different rules for the same activity and model range, the responsible business expert needs to resolve the relationship. The additional technical connection can record which sources were found; it does not thereby decide which conflicting rule the business should follow.

A skill in Claude Code contains instructions an agent can use for a particular task. It can also refer to supporting files or scripts. The agent performs the work using its tools. When we refer here to knowledge being retrieved “by the skill”, we mean retrieval performed by the agent under the skill’s instructions.

For our example, the skill would describe how to prepare a draft service handoff from the incoming report. It could instruct the agent to identify the assigned activity, retrieve its required knowledge sources and compare them with the case details. It should then prepare a case note, leave missing information unresolved and identify the source versions used.

A skill can contain several internal steps without each one appearing in the business process overview. A BPMN task describes work at the chosen level of detail; a collapsed subprocess refers to a procedure broken down into finer business activities. The existence of a skill alone therefore does not determine which representation is appropriate.

An understandable mapping is a useful starting point: which agent implementation handles this activity, which skill does it use and what result should it return? One skill may also serve several activities. In that case, it needs to be able to identify the business context in which it is currently being used.

For targeted retrieval, the agent needs an unambiguous reference to the process and activity. In our example, a general search for everything about the washing machine would be too broad. It needs the sources intended for “Prepare the service handoff” and, within them, the content applicable to this case. The process and activity references, and the model range where relevant, therefore need to be known when the assignment is made or the retrieval occurs.

MCP, the Model Context Protocol, can provide the technical connection between the AI application and the information made available. Which content belongs to the activity and how AI subsequently uses it follow from the business relationships and the retrieval implementation; the protocol itself does not determine them.

For this proposed design, the connection must expose the activity relationships and the necessary knowledge content. A link in the user interface alone is not sufficient. Check whether the configured access supports the required scope and which information it returns within the relevant permissions. Sources and versions need to be traceable so that the result can be reviewed later.

If the model identifier is unknown or no suitable source is accessible, that gap belongs in the draft. The agent should not select an arbitrary instruction merely to produce a complete note. The sample case below explicitly uses the fictional instruction SA-07 for WM-DEMO in version 2. Retrieval must establish which version actually applies to each case; a version number is not permanently fixed in the skill.

The result of the agent activity returns to Customer Service. The employee checks that the note reflects the available details correctly and identifies open questions. For our completed sample, assume that the customer has now provided the model identifier and described her observation more precisely. After review, the case note could look as follows; it is a constructed example, not the output of an agent test.

The case note distinguishes the customer’s observation from a technical finding. “Display stays dark” does not become “faulty electronics”. A request for service does not become a confirmed appointment either.

The actual handoff then takes place in the service application or through the execution mechanism configured for that purpose. A prepared note does not mean that technical service has already taken on the case. The overall process needs to make clear when the draft was reviewed, the case passed on and the next stage started. These states belong to the live case.

First check whether the case note is useful to the next person. Then examine the technical assistance along that same activity: did the agent receive the correct process reference, actually retrieve the appropriate knowledge source and apply its content correctly in the draft? A well-written note alone does not show whether it rests on the intended sources.

Use a complete case, a missing model identifier and conflicting sources. For each, examine the input, the retrieved sources and versions, and the returned draft. Missing details should lead to an appropriate follow-up question. An unresolved business conflict needs to remain visible, including what must be clarified before the work continues.

Whether the implementation uses an agent that works dynamically or a more prescribed sequence depends on the task. Anthropic distinguishes such agents from predefined workflows. A bounded language task within the service process does not automatically require AI to control the overall process as well.

Compare the new handling with the previous handoff, including questions, review and corrections. This shows whether the assistance helps the following process step or merely creates additional work elsewhere. The findings determine which parts of the procedure or knowledge foundation need adjustment.

Business owners need an understandable interface for reviewing knowledge pages and their use in the process. In PULSE, they can maintain the shared knowledge foundation and examine it in context. In the proposed design, the agent activity should reveal which instructions it needs and who can resolve an open business question. This also helps the people reviewing the drafts.

Ownership of the external skill concerns the agent’s procedure, its connections and the expected result. Business rules maintained centrally in PULSE do not then need to be copied into every skill. If a new instruction changes the required steps or output format, however, the implementation needs review as well. Central knowledge reduces duplicate content maintenance without removing every need for coordination.

• A process model with the assigned agent activity, its inputs, results and human handoff points.
• Business-validated knowledge relationships for the activity, including applicability and ownership.
• A documented external implementation with a skill and verified access to the required context.
• Example cases and specific review triggers, such as a new model range, updated instructions or a different handoff route.

For an unresolved item, establish who will address it and what will demonstrate that it has been resolved. Whether the current MCP connection supplies all required activity relationships and knowledge content is a specific technical question. Whether an instruction should govern intake is a business decision. Keeping these distinct makes the working session manageable even while some parts remain to be implemented.

The inventory helps record sources and their use at a process step. Use it for the knowledge relationships of your selected activity. Record the relationships in the inventory