An institutional agent needs a boundary before it needs a personality.
Define what it sees, what it proposes and who owns the decision.
Begin with a consequential question
A conversational answer can sound useful while drawing from incomplete or outdated information. Identify the workflow the agent serves: retrieving an approved guide, grouping observations or drafting a candidate improvement. Each task needs different sources and permissions. A public sales advisor should not pretend to see private field records. A departmental advisor should not generalize from one observation to the whole institution. Useful conversation begins with a clear knowledge boundary and an honest account of what remains unknown.
Keep evidence attached to the proposal
An improvement proposal should name the recurring friction, supporting observations, contradictory cases and the behavior expected to change. Show source dates and missing coverage. If observations are vague, the first recommendation may be better data collection rather than a new procedure. The agent can distinguish observed, inferred and hypothesis while helping a human reviewer understand the reasoning. A confident tone is not a substitute for evidence, and the number of inputs does not establish their quality.
Make the playbook living through review
Keep the approved version intact. Store a candidate revision separately with its rationale, proposed test and responsible owner. The owner accepts the test scope, reviews field use and decides whether the revision belongs in the next approved version. Preserve rejected proposals and lessons where appropriate. This is a learning system: it can explain how experience informed a change, while the institution retains decision authority and a traceable version history. The agent supports that path rather than silently rewriting it.
Free templates: record the decision and observe the friction