Separate the steps from the judgment
Consider a new customer inquiry. Copying its arrival time into a record is a predictable step. Reading a loosely written message and identifying the service requested involves interpretation. Deciding whether to make an unusual promise to the customer involves authority.
Those jobs do not need the same level of freedom. You can combine a regular integration, a bounded model call and a human approval without handing the entire process to an autonomous agent.
Anthropic distinguishes predefined workflows from agents that choose their own path and tools. It recommends predictable workflows for well-defined tasks, with more flexible agents where model-led decisions are necessary. Anthropic, Building effective agents, checked September 6, 2026.
Choose a workflow when the route is known
If a valid request always passes through the same checks, put those checks in code. Extract the fields you need, validate them and send the result to the next step. Reject or route cases that do not meet the rules.
This makes it easier to explain what happened. It also makes testing more concrete: given this input, the workflow should produce this output or this exception.
For many intake and document-handling jobs, that is enough. The model interprets the irregular part. It does not need to decide which systems it can access or whether to skip an approval.
Choose an agent when the path varies
Research can take a different route each time. One company has useful product documentation; another needs several public sources to establish its basic operation. An agent can choose the next retrieval step within the scope you give it.
Hermes supports reusable skills, memory and connections to external tools through MCP. Those capabilities can help a supervised research or operations assistant work across varied tasks. They still need configuration and access controls. Hermes skills documentation and MCP documentation, checked September 6, 2026.
Ask what happens when it is wrong
The important distinction is not how impressive the demonstration looks. It is whether an error is caught before it affects a customer, changes a record or spends money.
A useful design names the decision owner, limits the tools available and leaves a record of actions. A difficult task may justify an agent. It also justifies better testing and narrower permissions at the start.
Let the first build earn the next one
Choose a bounded workflow and agree on the acceptance checks. If it fails, fix it or stop. If it passes, review the exceptions before giving it more access.
Our custom-agent work begins with that boundary. Our back-office builds combine ordinary integrations with AI where the process calls for it.