Project change monitor

Use an agent when monitoring is an ongoing responsibility and new findings should remain in one conversation. The value comes from a precise significance rule, not from checking as frequently as possible.

Define the boundary

Create the agent, then turn off any inherited workspace connector it does not need to inspect. In its guidance, name:

• the approved sources and scope;
• the baseline or previous state to compare;
• what counts as material;
• who should be notified;
• what evidence every escalation must include; and
• what it must never change automatically.

For example: “Escalate only a changed delivery date, owner, blocked dependency, or decision that affects the committed milestone. Ignore formatting, routine comments, and unchanged status.”

Prove normal and edge cases

Test a material change, an irrelevant change, no change, and a failed source read. A source failure should be reported as incomplete monitoring, never as “no change.” Where operation controls are available, keep outbound messages and mutations ask-gated until the classifier is reliable; otherwise use read-only credentials and human delivery.

Choose the wake pattern

Use a Routine check-in wakeup when the agent should inspect its broad responsibility and decide what is due. Use a separately named wakeup when a concrete future check has its own instructions. Start with the slowest cadence that still meets the operational need and set a daily budget.

Manual and scheduled wakeup runs inherit the agent creator's personal no-training preference plus any organization requirement. If that effective policy is selected while enforcement is unavailable, d5s blocks the run before a model provider is called.

A defined check can also run as an independent automation, with the monitoring agent attached to its team. Leave Keep informed enabled for passive handovers, or explicitly choose Act on findings to wake the agent for failures or actionable outcomes. The agent can retrieve the referenced result under its current access; it does not share its full conversation with the check.

Periodically review the continuing conversation for repeated noise or stale assumptions. Refine the significance rule instead of increasing the model or check frequency first.