Anatomy of a change plan: what an agent shows you before it touches Jira
Every agent I run has to put its change into a plan first, short enough for someone to take in within a minute. If the plan cannot be read that fast, it gets rewritten before anything runs. Below is what a plan has to contain, shown on a small request: making one field required on one issue type.
Key takeaways
- A plan states what the agent found in the site before it proposes anything.
- Open questions are part of the plan. An agent that has none on a real site has not looked.
- Approval names a person. "Approved" without a name is a log entry nobody can follow up.
The five parts of a plan
- What discovery found. The current state of every object the request touches, read from the site, not assumed.
- The risk. What else uses those objects, and what breaks if the change is wrong.
- Open questions. Anything the request left ambiguous, asked before execution instead of guessed.
- The change, step by step. Each write the agent will make, in order.
- The approver. A named person, who can reject the plan or send it back with an answer.
JAMES writes its plans in that shape, and I hold agents I build for clients to the same one.
An example, on a small request
The request: "make Cost centre required on Purchase requests". The card and the change below are an illustration of the format, written for this note, not a record from a client site.
The interesting line is the baseline. A person asked for one field on one issue type, and the site answered that the setting is shared by three projects. An agent that skipped discovery would have made the field required everywhere.
Why the plan comes before the change
Reading first costs a few seconds of model time. Skipping it costs a morning of somebody working out why two other teams can no longer create tickets. The plan is also the record: once approved and executed, it is the answer to who changed what, and why. The loop itself is described on the method page, and how JAMES works shows it in a product.
Want this discipline around an agent in your own systems? The audit is where it starts.