Read time: 4 minutes 3 seconds
Think about the AI tools we use every day.
ChatGPT lives in its own tab. Claude has its own desktop app. Cursor sits alongside your IDE. Devin. Lindy. Always the same pattern.
The agents that promise to "work alongside you" tend to open in a private window, just you and the model. 1 human. 1 assistant.
This is a choice the entire industry made by default after ChatGPT, and almost nobody has questioned it.
Viktor is one of the only AI companies betting against it, and theyβve scaled to $15m ARR in just 10 weeks (!!) by doing so.
Today's breakdown is about what I'm callingΒ βmultiplayer AIβ.
β How Viktor went team-native by killing their own web app in two days
β Why 13,000+ teams hired Viktor as a colleague rather than a tool
β Why I think the rest of the AI stack is quietly going to follow.
I've been running Viktor inside the Strategy Breakdowns Slack workspace for a while now, and the thing that keeps surprising me isn't what it does when I ask.
It's what it does when I don't.
Shoutout to to the newest member of the Strategy Breakdowns team: Karen. Viktor is fired up.
Letβs get into it.
β Tom
P.S.Β Huge thanks to Antoni and the Viktor team for the candid behind-the-scenes context that made this one possible.
Don't SOC-block your best engineer
Your software needs to be compliant to win deals. But you also need your engineers focused on building your product β NOT pulling SOC 2 evidence.
The Vanta Agent is the sharpest GRC engineer you've never had to hire, working tirelessly across the platform to draft policies, complete questionnaires, and flag issues before they escalate.
Fast-moving companies like Ramp and Cursor use Vanta to get and stay compliant, simplify their audit process, and unblock deals β so teams can get back to building.
Own, store, and query your data in one place
PostHog is launching a managed DuckDB warehouse that'sΒ fully integrated with your product analytics.Β
Get direct database access, built-in SQL editor, modelling, and visualisation all-in-one, instead of a fragmented stack.Β
Thank you for supporting our sponsors, who keep this newsletter free.
Multiplayer AI
Chess Move
The what: A TLDR explanation of the strategy
The default architecture of modern AI is the personal copilot.
One human + one model + private window.
Viktor made the bet that the most useful AI is a part of a team. Their agent lives where the team already works, in Slack (and Microsoft Teams in ~2 weeks).
Sounds like a distribution decision but itβs also a positioning decision in disguise. When the data was showing that Slack is where users fell in love with Viktor, they killed their existing web app in 2 days.
13,000+ teams now treat Viktor as a colleague, with adoption that spreads *sideways* across the org instead of being pushed top-down.
What makes this multiplayer bet interesting is it challenges theΒ personal copilotΒ frame.
A private chatbot can't see (or be seen by) your team β So it can't behave like a teammate β So the product can never be more than a personal assistant.
Viktor's bet is that the next generation of AI products won't be the smartest single-player models. They'll be the ones structurally designed to hang with the team.
π‘
Strategy Playbook: Tabs have tools. Channels have colleagues.
Breakdown
The how: The strategic playbook boiled down to 3x key takeaways
1. Β Kill the personal harness
Almost every AI agent today is converging on the same UX: a chat tab in your browser.
The trouble with that surface area is that it's also the ceiling.
A personal chatbot harness can't see your team or act without being asked.
Always reactive. Never proactive.
Whatever the underlying model is capable of, the harness dictates what theΒ productΒ can and canβt do.
β A tab is a tool
β A channel has colleagues
When location dictates function, distribution = positioning
"If you ask Viktor in Slack to build a Stripe dashboard, it goes away, comes back three minutes later with the actual dashboard, and that's magical. If you ask the same thing in a web app, you sit watching a spinner for three minutes wondering if it's broken. Same product, same answer, opposite experience.
The web app put us in the ChatGPT frame, where users expect sub-second responses. Slack put us in the colleague frame, where three minutes is fine.
We killed the web app the week we figured out the surface was choosing the expectation.β
1) The addressable surface goes from "people willing to add another AI tool" to "people already inside Slack" (200m+ people)
2) A whole class of product behaviours becomes possible the moment you're in a shared channel:
In-channel context
Multi-stakeholder workflows
Work that other teammates can see
We never told Viktor we run 7-day and 14-day ad performance reports. It just spotted the pattern from prior weeks, sent a reminder, and even offered to pull the numbers independently going forward. Viktor now runs our reporting on an ongoing basis.
None of that is available to a strictly personal harness, no matter how good the model is.
Where your tool lives tells the user what you are.
2. Make adoption a side effect
Word of mouth depends on someone remembering to mention your tool. The personal-copilot model creates little room for organic viral loops because nobody else in your company sees you using it.
But when the agent is visible in shared channels, adoption becomes a side effect of just using the product.
βHalf of paying Viktor workspaces have it active in 5 or more distinct Slack surfaces within 30 days. The top 10% have it in 19 or more.
That's not a procurement decision spreading top-down. That's a colleague being invited into more rooms.β
Viktor comes across as a "good vibes" colleague with professional boundaries, the kind of teammate who won't throw someone under the bus in a public channel.
βIn a shared channel, voice is what tells the team whether your agent is a tool or a teammate. So we built personality as a deliberate product decision, not a brand afterthought.
Viktor has opinions, defends them, and pushes back when something doesn't sit right. That's the version that earns its way into more channels.β
3 important decisions stack here:
Responses are public-by-default in shared channels, so its work is visible
Viktor has a recognisable personality rather than your everyday AI assistant voice
It shows up in small social moments inside a team's Slack (celebrating wins, spotting patterns, identifying opportunities)
A team installs Viktor, and within days, other teammates start tagging it, without being onboarded.
The agent gets pulled across teams and into new channels, the same way a useful colleague does.
Proactive colleague, not reactive prompt-receiver
3. A teammate, not a chatbot
Every chatbot waits to be prompted, which puts an important burden users.
β Users have toΒ rememberΒ to use the AI.
The copy-paste-prompt loop most people are stuck in is a mental ceiling on what AI can be at work.
But in the real world, a great teammate doesn't wait to be tagged.
They notice things.
They prep the doc before the meeting.
They surface the report before someone asks for it.
They celebrate a win without ever needing a βtriggerβ.
Theyβre just βon itβ.
That's the move from AI-as-utility to AI-as-colleague, and it's more aboutΒ initiative than intelligence.
No one asked Viktor for this post-send update. It just took initiative.
Inside the SB-Lab workspace, Viktor now runs 3 crons, 27 skills, 7 integrations (Slack, Notion, beehiiv, Google Analytics, GitHub, Attio, Coworker Slack), and picks between 3 models per task.
A proactive AI agent builds context every time it acts β which makes the next action better β which earns it the right to be in more channels β which gives it more users and more context.
The flywheel only turns if the agent is allowed toΒ startΒ the conversation.
βThe hardest engineering problem wasn't the agent itself. It was the surface that lets the agent volunteer for work without being annoying.
Most of what makes Viktor feel proactive is plumbing: pattern detection, permission models, channel etiquette. The model is the easy part.β
The chatbot mental model is single-player.
The colleague mental model is multiplayer.
Rabbit Hole
The where: 3x high-signal resources to learn more
[1 minute setup]
We've had Viktor running in the SB Slack for a few weeks now, and we wonβt be going back.
If you want to feel the difference between a chatbot and a colleague, this is the best way to get started.
$100 in Viktor credits on us, no card required. Drop it into your team's Slack and see what happens in week 1.
[2 minute read]
Co-founder Fryderyk's launch thread - the moment the multiplayer bet took over tech Twitter. Worth reading not just for the product pitch but for how the team framed the bet from day one.
'AI coworker that lives in Slack' - the positioning baked in from the start
3,000+ tool integrations as the proof point
The clearest signal of what Viktor was building toward, before anyone outside the team had a name for it
[3 minute read]
One early Viktor user racked up aΒ $5,000/month billΒ from a single cron job firing every 5 minutes.
One of infinite failure modes that only emerge when your agent has access to 3,000 integrations and 200+ tools per user. A field report from the messy frontier of agent design.
Whenever you're ready, there are 3 ways we can help you:
Our flagship course on how to use free internet data to make better strategic decisions. Contains 5 years of strategy expertise, proven methods, and actionable tactics to accelerate your career with modern-day strategy skills.
We have a growing audience of 125,000+ strategists from top companies like Google, Meta, Atlassian, Stripe, and Netflix. Apply to feature your business in front of Strategy Breakdowns readers.
One of the most common questions we get asked is: βWhat tools do you use to run Strategy Breakdowns?β So, weβve open-sourced our tech stack to give you an inside-look at exactly what tools weβre using to power each corner of this operation.