Skip to content

Start with One Real Job

The fastest way to understand Codeer is to bring one recurring situation where a knowledgeable person still needs to judge the result.

Do not begin by configuring every feature. Begin by writing a short job brief:

  • Situation: What has just happened, and who needs help?
  • Result: What should the Agent help that person accomplish?
  • Quality: What would make the result correct and useful?
  • Boundary: What must the Agent avoid, clarify, or return to a person?

For example, the job might be to answer an approved course question, guide a teacher toward a practical next step, or review a draft against a professional standard. Choose one situation you recognize from real work.

What you will complete

At the end of this guide, you will have:

  • A first Agent draft for one defined job
  • One real scenario that demonstrates the job, plus a few nearby boundaries
  • Reusable must-pass cases with checkable expectations
  • A decision to publish a controlled scope or keep improving before launch
  • A way to handle and learn from new conversations after launch

This guide does not promise that one passing test proves business, service, or learning outcomes. It gives you evidence about the Agent behavior you inspected.

The release gate belongs to your team

Test Suite keeps cases and results, and Agent version history keeps working versions. Codeer does not automatically block publishing when a required case fails, or store the release approver and stop conditions as a release record. Before publishing, your team must choose the must-pass set, review the evidence, and record the decision described in Publish Only the Verified Scope.

Who decides what

Role Main decision in this guide
Expert or owner What good looks like, what is unsafe, and which scope can be published
Operator Which conversations need action and which new situations should become cases
Admin or implementer Which workspace, sources, Channels, permissions, and tools the Agent can use

One person can hold more than one role. Keep the decisions separate even when the person is the same.

1Choose the workspace and job
↓
2Create a first draft
↓
3Try the real situation
↓
4Keep must-pass behavior
↓
5Publish a controlled scope
↓
6Handle, learn, and verify again

Step 1: Choose the Workspace and Job

Use an existing workspace when the Agent should share the same customers, knowledge, operators, and memory scope. Create a separate workspace when those boundaries should stay isolated.

Write the job brief before opening the Agent Editor. If you cannot describe the result and boundary yet, use one real situation to make them concrete.

→ Set Up Your Workspace

Step 2: Create a First Draft

Open AI agents in the workspace and click New.

  • Use Use a template when a Template matches the work. Answer the setup questions, generate the draft, and review every field before creating the Agent.
  • Use Start blank when no Template matches. Describe the job, quality bar, and boundary in plain language; Copilot can help turn them into a first draft.

Both paths return to the same Agent Editor and the same verification steps.

→ Use a Template

→ Build the First Version

Step 3: Try the Real Situation

Use Live Test to run the situation from the job brief. Then try a few nearby boundaries: missing information, a request outside scope, a risky promise, or a point where a person should take over.

At this stage, you are inspecting the first behavior, not proving the Agent is ready for every user.

→ Build and Debug the First Draft

Step 4: Keep Must-Pass Behavior

Save the situations that must remain correct as reusable cases in Test Suite. Give each case a Standard that another reviewer can use without guessing.

For a first visible check, one real scenario and a few nearby boundaries are enough to expose the design. Before a controlled pilot, expand the set according to frequency, consequence, tool use, and handoff risk. There is no universal case count.

→ Verified Scenarios

Step 5: Publish a Controlled Scope

Publish only after the expert or owner has reviewed the must-pass behavior and the fallback for unverified work. Keep the first audience small enough that the operator can observe and respond.

→ Publish Only the Verified Scope

Step 6: Handle, Learn, and Verify Again

Use Conversations to see what actually happened. Respond when a person is needed, save important new judgment as a case, make the smallest relevant change, and rerun the affected cases before publishing another version.

→ Review Conversations and Improve

→ Reply to Conversations

Next steps