1 / 15
↑↓ navigate  P presenter
Seven Shades of Agents

Autonoma Agenter – vad är det och vågar vi?

A field report on autonomous agents, real systems, and trust.

Johan Wallquist · Partner Solution Architect · Microsoft

Scroll or press ↓ to begin
Seven Shades
of Agents
1
Ask & Answer
2
Draft Partner
3
In-Context
4
Task Execution
5
Team Direction
6
System Op.
7
Dark Factory
Level 1

Ask & Answer

You ask questions. AI answers. Like a smart search engine — fast, broad, but you do all the work.

Core insight: AI helps you find information. You decide what to do with it.

Example: ChatGPT, Bing Chat

Level 2

Draft Partner

AI generates first drafts you refine. Emails, documents, summaries — faster starts, your judgment on the finish.

Core insight: AI creates starting points. You shape the result.

Example: Copilot in Word, email drafts

Level 3

In-Context Assist

AI works inside your tools. It knows your calendar, your files, your data. It doesn’t just answer — it understands your context.

Core insight: AI is embedded in your workflow. It sees what you see.

Example: M365 Copilot in Teams/Excel/Outlook, Copilot Cowork

Level 4

Task Execution

You describe a task in plain language. AI plans the steps, executes them, and checks its own work. You review the result.

Core insight: AI doesn’t just suggest — it acts. Open-source coding agents like OpenClaw prove the pattern works at scale.

Example: Copilot CLI, Cursor, OpenClaw-type coding agents

Level 5

Team Direction

You direct a team of specialist agents. Each has its own expertise, memory, and role. They persist across sessions and learn your project.

Core insight: You’re not writing prompts — you’re directing a team.

Example: Microsoft Agent Framework, CrewAI, Squad

Level 6

System Operation

Agents run autonomously — discovering work, triaging priorities, executing approved tasks. Humans step in at judgment points. The system operates while you sleep.

Core insight: Humans decide what matters. Agents handle everything else.

Example: Autonomous DevOps pipelines, AI-powered content systems, self-healing infrastructure

Level 7

Dark Factory

Fully autonomous. You set the vision and values. AI defines the architecture, builds the system, operates it, and improves it. Software that builds itself.

Core insight: The destination on the horizon. Not here yet — but look how close we got.

Example: The emerging frontier

Level 1 / 7
Interactive

Where Are You? Where Is Your Organization?

Before I show you the result — let’s see where this room is at.

✋
You personally
What level are you at today?
🏢
Your organization
Where is your company?
🚀
Sweden's coolest startups
Where are they?
The shades: 1 Ask & Answer · 2 Draft Partner · 3 In-Context Assist · 4 Task Execution · 5 Team Direction · 6 System Operation · 7 Dark Factory
Show of hands — hold up the number of fingers matching your shade. 1 through 7.
Level 5 — Let me prove it

We could talk about autonomous agents.
I’d rather talk to them.

Tip: Watch what happens in the empty folder on the left side of the screen.
Level 6 — Here’s what I built

Same conversation. Two weeks later.

I initiated the build. Reviewed. Iterated. Kept going.

🌐
Live site — 10 categories, full-text search, community access
📡
14 sources scanned daily — AI discovers content overnight
🤖
Agents research, validate, assign, format, ship — approved content goes live in minutes
👥
Domain experts curate — they decide, agents execute
“Built by one person. Who can’t code.”
Level 7 — The horizon

Dark Factory

Fully autonomous software. Systems that build and operate themselves. We’re not there yet.

“But look how close levels 5 and 6 already got us. The path is clear.”

Why now

AI Was a Great Advisor.
Now It’s Also a Great Operator.

  • The threshold: Coding agents crossed the quality line — AI can now build and ship real changes to real systems. Everything is code; when AI masters code, everyone gets access.
  • Reasoning: Multi-step planning, dependency handling, complex workflows
  • Tool use: MCP, custom skills, specialist agents, extensible toolboxes
  • Self-correction: Agents check, iterate, and improve — in loops, not one-shots
🧠 Understand
→
📋 Plan
→
🔧 Act
→
✅ Check
→
🔄 Iterate
Wrapped in memory · guardrails · human review
Start simple

The Building Blocks

Start with a conversation. Add tools. Then add a team.

💬
Copilot CLI
The foundation. You describe what you want. It plans, acts, checks.
→ Level 4
🔌
+ Tools
MCP servers, custom skills, specialist agents, pre-built toolboxes connect Copilot to any system.
→ Level 5
👥
+ Team
Hire specialists. Each with their own expertise, memory, and task list.
→ Level 6
Proof

By the Numbers

2
evenings to first version
8
specialist agents
14
sources scanned daily
75+
curated entries
~15
min/day to operate
1
non-developer
The trust question

Vågar Vi?

What happened when I trusted too much.

🔥 Day Three — One agent. Every task. I asked it to add a link. It added the link — but also “improved” three existing entries I never asked it to touch. The output looked professional. I approved it. The site broke.

“Agent confidence ≠ agent correctness.”

The answer

The Trust Spectrum

Not more control. Not less control. The right control.

🔒 Check Everything
Safe but slow. Nothing ships without full review.
✅ Trust but Verify
Humans at judgment points. Agents handle the rest.
⚡ Let It Run
Fast and autonomous. Monitor, don’t micromanage.
  • Specialist agents → narrow scope, fewer surprises
  • Validation layers → agents verify their own output — and each other
  • Human checkpoints → at the decisions that matter most
  • Full audit trail → every action logged, every change reversible
What makes it safe

Guardrails That Make It Work

🔒 Security & Governance
  • Human review before production
  • No secrets or sensitive data in prompts
  • Full audit trail · scope limits · rollback
🤖 Responsible AI
  • Agents can be confidently wrong — verify
  • Document what’s AI-generated
  • Start small, earn trust, expand scope

“Autonomy without guardrails is just chaos with better PR.”

Three takeaways

What I Learned

1
We’re already there.
The tools work — well enough to build and operate real systems with human oversight. Not someday. Now.
2
Trust is earned, not assumed.
Start small. Let agents prove themselves. Expand scope as confidence grows. For AI — just like for people.
3
The teams that direct agents will move faster.
This is becoming a competitive advantage — and the window is open now.
Invitation

What Level Are You At?

The people who used to be blocked from building were not blocked by lack of ideas. They were blocked by access — not by ambition.

That access is here now. Vågar du?

Get started

Resources

Get this presentation, plus additional links, tools, courses — and connect with me — in one place:

QR code
Scan to open

AI was a great advisor. Now it is also a great operator.
What will you operate?