Build a Governed Agent

The course

About this course

The follow-on to the flagship. A small group takes one process the roadmap chose, usually a reconciliation, a load validation, a missing-mapping proposal, or the morning digest from the close calendar, and builds it into an agent a second person can run and reproduce. The deterministic parts become code with tests. The AI parts read exceptions and draft explanations. Nothing lands in a system of record without an approval, and every run is logged. Systems people work in Claude Code; process owners work with a skill file and a script. Both leave with the same documented, governed agent.

Who it is for

Process owners and finance systems, reporting, and data people: the ones who will own the first two items on the roadmap

Format

Hands-on half day for a small group. One process from the roadmap becomes a scheduled, logged agent with an approval gate.

What the team leaves with

Outcomes

• One documented, governed agent: connection, rules as code, AI for the exceptions and the narrative, an approval gate, a log, a schedule.
• A validation script for a load file, or a reconciliation between two exports that reports what changed, written and tested.
• A rules file and a README that let a second person run it and a third person audit it.
• The judgement to know which roadmap items belong in code, which belong to the tool, and which stay with a person.

Outline · 2.5 hours

How the session runs.

Usually 9:30 to 12:00 or 2:00 to 4:30. Other times are fine.

1. 0115 min

   Pick the process and draw the boundary

   • One roadmap item. Inputs, outputs, who approves, where it logs.
   • The read-only boundary: what the agent may touch and what it may only propose.
2. 0230 min

   A connection instead of an upload

   • Pointing the tool at a folder, an export, or a query instead of pasting.
   • The project rules file that makes it behave like a colleague who knows your platform.
   • Where it is reliable and where it is confidently wrong.
3. 0335 min

   Rules as code

   • The deterministic core: sums, joins, as-of logic, thresholds, written as a script with tests.
   • The AI layer on top: read the exceptions, classify timing versus error, draft the note.
   • Generating the test cases nobody had time to write.
4. 0440 min

   The approval gate, the log, the schedule

   • Output lands in a review folder or an email, never in the ledger.
   • Every run writes what it did and why. The approval is the trigger for anything that posts.
   • Scheduling it, and the morning digest that says what ran and what failed.
5. 0530 min

   Can someone else run it

   • Swap seats. A second person runs the agent from the README.
   • What to fix so they can. Conventions: where the rules live, what gets committed.
   • The second roadmap item, scoped.

Bring

What to bring

• A laptop with a terminal and a Claude login your organization allows for work. Setup instructions come ahead of the course.
• The roadmap from the flagship course, and sanitized copies of the inputs for the process you are automating.

If IT blocks the tool, tell us on the call and we plan around it.