← Operator Hour
Season 1 · Episode 6 · Public recap

The sidekick was an agent. Now build yours.

The finale opened up the agent used throughout the season, then rebuilt the same pattern around a different team task.

Recap

The internal support agent was called Otto. It was a Microsoft-only cohort sidekick with a narrow job: answer questions from approved season material and help participants find the right exercise. Its conversations, identity, tenant configuration, and deployment remain private.

For the public version, we changed the role and replaced the knowledge with fictional onboarding material. The result was a second agent doing a different job.

The code mattered less than its boundary. Answer from these files. Admit when the answer is unclear, and send people back to a human when the material runs out.

The modelThe persona sets the behaviour. Knowledge defines the limits. Change the files and you change what the agent can do.

The original internal agent and its deployment details remain private. We rebuilt the reusable architecture as a clean public blueprint using synthetic examples.

Operator move: pick one of the questions your team answers every week. Put the approved answers in a folder, then give an agent a narrow job and a clear stopping point.

Take it home

Start with one repeated question

Choose a small body of knowledge that is safe to share with the intended users. Be clear about what the agent may answer and when a human needs to take over.

Do the exercise

Turn a safe folder into a narrow assistant

Use the public Assistant Blueprint. It contains synthetic policy, memory, and workflow examples without Otto's internal code or Microsoft configuration.

30 to 45 minutesPublic synthetic starter
  1. Fork or clone Executive Assistant Blueprint.
  2. Read the boundary policy before adding knowledge.
  3. Create one synthetic FAQ with five questions for a fictional team.
  4. Use the prompt below to adapt the skill, then run the included public-safety check.
Narrow assistant prompt 
Read this repository before changing files.

I want a narrow assistant for a fictional team's onboarding FAQ.

First:
1. Explain the boundary policy in plain language.
2. Show which committed files are safe templates and which real values must stay local.
3. Propose the smallest change that proves the assistant can answer from five synthetic FAQ entries.

Wait for approval.

Then:
- Add only synthetic knowledge
- Update the portable skill for this narrow job
- Make uncertainty and escalation behaviour explicit
- Run the repository safety checks
- Show me what remains mocked or platform-specific

Do not add connectors, credentials, real people, customer data, or enabled schedules.
You are done when: the assistant answers from the synthetic FAQ, refuses an unsupported question, and the repository safety checks pass.
Keep it safe: a public starter proves the pattern. Real knowledge, connectors, memory, and schedules belong in an approved private environment.
Resources
🎞️

Public slide deck

The narrow-agent model, Otto boundary, exercise, and public blueprint.

open slides 🎬

Full recording

Microsoft employees can watch the season agent taken apart and rebuilt after corporate sign-in.

Microsoft employees 🤖

Assistant Blueprint

A public, synthetic starter covering boundaries, memory, knowledge, and scheduled work.

public repo 🧭

Copilot Overview

A smaller public example showing how repeatable agent work can be packaged as an installable capability.

public repo

Still not a developer. I just gave one repeated team task a narrow agent and a clear boundary.