Runs on your own infrastructure if you want it to. Your data never leaves.
Current agent team
Chief of Staff
Personal Assistant
Meeting Analyst
Product Manager
PMM
Customer Success
Copywriter
Sales Rep
Event Promoter
Presentation Studio
Product Designer
Developer
Code Reviewer
Researcher
SEO team
Investor Relations
Maskin · 04 · Poll
Poll #1 · Luma
Have you experimented with AI agents running in loops?
A
Yes, running today
B
Yes, tried once
C
No, not yet
D
What's a loop?
Whichever bucket you're in — you're about to see three real ones I run every week, before we define anything.
Maskin · 05 · Real loops, mine
Three loops I run today · not concepts
Before I define anything — three loops I actually run.
LOOP 01 · MASKIN AGENT
Meeting Analyst
Every customer call turns into tagged insights, in the workspace. No notes-writing after the call.
LOOP 02 · MASKIN AGENT
Presentation Studio
Every event brief becomes a designed deck on the event object — including this one.
LOOP 03 · MY OWN
SEO team
One loop, 30 SEO posts a week. My personal newsletter runs on it.
All three: work happens without me driving. That's the shape we're going to build on.
Maskin · 06 · Wedge
The debunk
A chat is not a loop.
A CHAT
You → prompt Model → answer You → prompt again Model → answer You are the loop.
A LOOP
Goal → agent runs Output → coach reviews Signal → next cycle Cycle N+1 > N The agent closes the cycle.
Maskin · 07 · Cost of status quo
The tax you're already paying
58%
of a knowledge worker's week disappears into repetitive, low-judgment work — coordination, chasing, writing up.
Source: Asana — Anatomy of Work · 2024 · pending Researcher final pass
SUPPORTING
7.5 hrs / person / week
just searching for information that already exists in the company.
McKinsey Global Institute
WHERE IT LEAKS
Meeting notes not written. Backlog un-groomed. Insights forgotten. Follow-ups skipped.
Every hour here is an hour your team isn't doing what only they can do.
Maskin · 08 · What smart work looks like
The shift
What would your week look like if 10 agents ran 10 loops for you?
85%
of 20,000 companies deploying agents report clear productivity gains — average 2.5 hrs saved per employee per day.
HubSpot AI Adoption Report · Q1 2026 · pending Researcher final pass
You define the outcome. You review the work. The middle is agents in loops. That's what Maskin already runs on.
Maskin · 09 · How we talk to AI
How we've talked to AI has changed
What changed isn't the model. It's where you sit.
2022
Chat
Human is the loop
→
2024
Tools / Copilots
Human still driving
→
2026
Agents in loops
Human on the outside
Every step outward is one more piece of work that doesn't need you in the middle.
Maskin · 10 · The primitive
What is a loop
A loop.
The primitive is the loop, not the agent count. One agent or many specialized agents — still one loop.
Maskin · 11 · One or many
Single-agent — or multi-agent
Same loop. One agent, or a team.
SINGLE-AGENT · EXAMPLE
Daily competitor price check
One trigger. One agent. One output. Runs every morning without you.
MULTI-AGENT · EXAMPLE
Bug triage → assign → resolve
Specialized agents, handing off inside the loop. Still one loop, still closed.
The primitive is the loop, not the agent count.
Maskin · 12 · Anatomy
What is an agent
An agent.
Maskin · 13 · Skill ≠ agent
Don't confuse them
A skill is one ingredient. An agent is the whole recipe.
SKILL (bare Claude / OpenAI)
One-shot · stateless · you invoke it
Great at one task, forgets everything the moment it's done. Someone has to call it. Someone has to decide when.
AGENT (Maskin)
Persistent · owns an outcome · decides when to act
Named. Has a driver. Sits in a loop. Wakes up on a trigger, does the work, compounds what it learns.
Maskin · 14 · Hiring criteria
What makes a good agent
You're hiring. Not configuring.
01
Clear goalNot "help with things." A named outcome you can point at.
02
End-to-end responsibilityOne owner from trigger to close. No handoff into the void.
03
Right tools + skillsReach into the systems it needs, know how to use them.
04
A coach reviewing every cycleSomeone (or something) closes the loop by grading the work.
Maskin · 15 · Your turn
Design your own loop
Pick a loop from your work.
Hold these four questions in your head while I show you two live. No answers required — just carry them into the demo.
Question 01
What triggers it?
A schedule · an event · a message?
Question 02
What closes it?
How do you know the cycle is done?
Question 03
Who's inside?
One agent, or a team of specialized ones?
Question 04
Where does state live?
Whose memory holds what worked?
Part I · Live in the app
Let me show you.
Fresh workspace. Two loops built live from copy-paste prompts. In the app, on stage.
Loop A · live demo
Office DJ weather → Spotify.
TRIGGER · 08:00 daily — every morning, Copenhagen time.
AGENT · DJ — reads weather, picks a vibe, pulls a Spotify playlist, posts to #demo.
CLOSE · 08:15 — playlist dropped by 08:15.
Set up a loop for me. Every morning at 8am Copenhagen time,
check the weather and pick a Spotify playlist that matches the
vibe — mellow if it's rainy, upbeat if it's sunny, that kind
of thing. Post the playlist link to the #demo channel with one
line explaining why. Have it done by 8:15.
◐ Alt+Tab to Maskin — paste prompt A1 in a fresh workspace
Maskin · 18 · In vs on
Where do you sit?
Human in the loop · Human on the loop.
Human IN the loop
Bottleneck. Throughput capped. Nothing compounds.
Human ON the loop
You own outcome + review. Agent owns the cycle.
Same loop · in-app
Where are YOU in the weather loop?
Walk the UI. Show what "in" would look like (approve every playlist). Then flip to "on" (only review outcome). No new prompt.
◐ Alt+Tab to Maskin — same loop, same workspace
Maskin · 20 · Open loop
Open — where most people stop
An open loop.
Work gets done. Nothing learns. Yesterday's playlist was terrible? Tomorrow's is just as bad.
Loop A · close it, live
Open → CLOSED.
The closing element is you. Your 👍 / 👎 in #demo is the signal.
Say first, then paste
"Everyone — join #demo now. Your reactions become the loop's training signal."
Update the Office DJ loop. After each morning post, watch #demo
for thumbs-up and thumbs-down reactions for 24 hours, and
remember which playlists worked for which weather. Next morning,
pick based on what got a good reaction last time in similar
weather.
◐ Join #demo · react live · cycle N+1 > N
Maskin · 22 · Good loops
What makes a good loop
Good loops look like good hires.
01
Clear outcomeNot "make a playlist" — "match the office vibe."
02
A named close conditionWhen is this loop done? A signal you can point at.
03
End-to-end responsibilityOne owner from trigger to close, no dropped batons.
04
A coach reviewing every cycleLoop stays honest because someone (or something) grades it.
Loop B · Stage 1
Meeting transcript → insights.
TRIGGER — a new customer meeting transcript lands.
AGENT · Meeting Analyst — reads it, pulls out every bug, feature request, feedback.
CLOSE — each finding stored as an insight, tagged with customer + meeting.
Whenever a new customer meeting transcript lands in the workspace,
have Meeting Analyst read it and pull out every bug, feature
request, and piece of feedback. Store each one as an insight,
tagged with the customer name and the meeting it came from.
◐ Alt+Tab to Maskin — paste prompt B1
Loop B · Stage 2 · + PM
+ Product Manager prioritizes.
NEW AGENT · Product Manager (Maskin) — clusters insights by theme.
SCORE — how many distinct customers raised it.
OUTPUT — a live prioritized backlog, updated every meeting.
Add Product Manager (Maskin) to the customer-meeting loop. When
new insights come in from Meeting Analyst, have PM review
everything from the last 30 days, cluster by theme, rank each
theme by how many different customers mentioned it, and update
the backlog with the new priority order.
◐ Alt+Tab to Maskin — paste prompt B2 · now multi-agent
Loop B · Stage 3 · + Coach
+ Meeting Coach for me, personally.
PARALLEL LOOP · same trigger — the transcript feeds two loops now.
OUTPUT · < 1 hour — short coaching note. Patterns compound over meetings.
Add a second loop off the same customer meeting trigger — Meeting
Coach for me. Analyze the transcript for how much I talked,
whether I asked open or closed questions, moments I missed a
follow-up, and the energy in the room. Send me a short coaching
note within an hour of the meeting. Track patterns over time
so each meeting builds on the last.
◐ Same trigger → two loops · one for the product, one for me
Maskin · 26 · Recap
Recap
Different shapes. One primitive.
LOOP A · WEATHER
1 trigger · 1 agent · 1 output
Silly, small, closed by audience thumbs-up in #demo.
LOOP B · MEETINGS (3 STAGES)
1 trigger · 3 agents · 2 parallel loops
Insights → prioritized backlog + parallel coaching loop for me.
Define outcome → agent(s) run → coach closes. That's the whole game.
Maskin · 27 · Poll #2
Poll #2 · Luma
What's stopping your team from closing loops?
A
Trust in the output
B
No way to measure quality
C
Model changes break it
D
Don't know where to start
Maskin · 28 · Evals as defence
Quality · the objection we always get
"As hallucination picked up, the quality wasn't really there."
SEO attendee · CPH meetup · 2026-08-25
Behaviour contracts + regression tests on the loop, not on the model. Any model can be swapped in.
Maskin · 29 · Two new jobs
Your two new jobs
Everything else delegates.
DEFINE
Name the outcome. Draft the loop.
REVIEW
Grade the work. Signal what to change.
Maskin · 30 · Vision · next
What comes after loops
Loops that hire loops.
The ask · book a Loop Sprint
AI-Native Loop Sprint.
Live in 2 weeks — built with you, not handed to you. 3 loops shipped, team trained, tools wired.
€7,500fixed
Setup only. Platform usage billed at €1.20/M tokens — or self-host free (Apache 2.0).
Duration
2 weeks · kickoff day 1, live day 14
What ships
3 production loops · Slack + CRM wired · team trained