An AI agent job card is a one-page assignment for a builder: trigger, approved sources, output, reviewer, forbidden actions and the failure path. Write the card before you connect any records. Then test whether the agent returns the draft, flags gaps and stops for review instead of deciding or sending.
What goes on an AI agent job card?
The card defines one repeatable sequence. Which file starts the job. Which fields the agent reads. Which reference it may use. Where the draft waits for review. The interface does not matter: a chat window and an automated workflow need the same written boundaries.
Compare asking your office coordinator a question with assigning her the weekly job of assembling superintendent notes. The second needs a source location, a due point, a format, an exception path and an approver. An agent needs the same.
Which fields belong on the card?
| Field | Builder example | Control |
|---|---|---|
| Trigger | A new warranty form lands in the service queue | Only approved forms start the job |
| Sources | The request, the project record and the current warranty workflow | Superseded files stay out |
| Output | An intake summary with missing information flagged | No coverage decision appears |
| Reviewer | Whoever owns service intake | The draft cannot skip the queue |
| Failure path | Unknown address or unreadable attachment | Stop and ask for a person |
Every assignment in the AI for builders library fits on a card like this. An agent without one is a chat window with ambitions.
Which agent roles fit a building operation?
Give agents bounded office work, never executive authority. A research agent collects public permitting pages and returns the links. A drafting agent turns approved field notes into a buyer update. A checking agent compares the spec sheet with the selections log and flags labels that do not match.
A building agent can format approved content into a website page or an internal form, with acceptance checks. A coordination agent can move a reviewed draft into the right queue. None of them picks a trade, promises a completion date or releases a draw.
How do several agents work as a fleet?
A fleet splits one workflow into visible roles. One agent gathers sources. One drafts. One checks the draft against the source and the fixed rules. At AgAI Labs we run a fleet of Claude and Codex AI agents on exactly that split for research, writing, building and checking. People set the direction and review every output.
The split makes errors easy to find. If a draft names a plan feature that is not in the approved plan set, the checking role flags it. The reviewer can then see whether the source was wrong, the drafting instruction was loose or the check missed it. One opaque tool that does everything leaves no trail.
Which job do you give an agent first?
Pick work that repeats, has a stable source and already gets knowledgeable review. Meeting-note cleanup qualifies because the recording and the attendees are the reference. A permit register qualifies once the jurisdiction portal and the project identifier are known. A selections exception report qualifies when the selections log is current.
- Name one output. Ask for an intake summary, not "handle warranty."
- Name the sources. The current plan revision, the log or the submitted form. Nothing else.
- List forbidden decisions. The agent cannot approve, diagnose, commit or send.
- Write test cases. One complete record, one missing field, one pair of conflicting sources.
- Assign the reviewer. The person who understands the underlying work.
- Run it beside the existing process. Compare the agent's draft with what the team would have produced.
- Record corrections. Turn repeat reviewer fixes into tighter instructions or cleaner sources.
The AI operations stage starts with one narrow assignment and a manual fallback, not a software catalog.
Where do agent assignments fail on real projects?
They fail when the source is not controlled. A project folder with three plan revisions gives the agent three plausible answers. A selection named one way in the allowance schedule, another in the design notes and a third in the purchase record produces false exceptions. An unlabeled superintendent photo cannot prove which house or room it shows.
They also fail when approval is implied instead of designed. A draft dropped straight into an outbound mailbox looks ready even when nobody read it. A summary pasted into the project record gets mistaken for a field observation. Working files need visible draft status. Approved outputs need a named owner.
What stays with people?
People interpret contracts, approve scope, decide coverage, commit schedules, direct safety work, authorize payment and settle field conflicts. Agents organize the information around those decisions. They do not hold the site context, the authority or the accountability.
The article on what AI should not do draws the boundary list in detail. The human review checklist turns it into an approval routine.
How do you know an agent job is worth keeping?
Judge the output and the burden of producing it. Track what the reviewer corrected, which missing inputs stopped a run and whether the draft arrived where the team could use it. Speed of text on screen tells you nothing.
A good assignment removes blank-page work and leaves judgment where it was. The reviewer sees source references, open questions and the status of each required field. If the team has to reconstruct the reasoning every run, the assignment is not ready.
Once the card passes its tests, it becomes the runbook: trigger, sources, output, reviewer, prohibited actions, failure path and manual method. Keep the accounts, files and approved outputs in your own control. The AI for home builders guide covers the wider operating model for source control and review.
Which job card questions come up first?
Does an AI agent work without instructions?
An AI agent needs a defined task, approved sources, an output format, boundaries and a reviewer. Broad instructions produce broad, unreliable work.
Can one agent run the whole office process?
One agent running the whole office process holds hidden authority. Split the work into preparation, checking and human approval so failures stay visible.
Should the agent have access to every project file?
Give an AI agent only the current material its assignment needs. Exclude credentials, unrelated buyer data and superseded records.
What happens when sources conflict?
When an agent's sources conflict, the agent stops the affected work, names both sources and routes the conflict to the reviewer. It never picks the convenient answer.