AI intel digest
Viktor: AI Coworker That Lives in Slack — Fryderyk Wiatrowski
Viktor is an AI employee that operates entirely inside Slack with no web UI, launched in February 2026 by Fryderyk Wiatr
Executive summary
Viktor is an AI employee that operates entirely inside Slack with no web UI, launched in February 2026 by Fryderyk Wiatrowski and team. The talk explains why scaling from a personal agent to a company-wide agent breaks on memory isolation, Slack's complex interaction surface (threads, DMs, edits, reactions), and model personality sensitivity. The core thesis: a company agent must live where the team lives, inherit shared integrations once rather than requiring every user to connect their own, and earn proactivity gradually to avoid security backlash.
Signal points
- 1
Scaling from 1-user to 100-user agents breaks memory isolation in ways that are 100x worse, not just linearly worse
- 2
Slack's complexity (threads, DMs, edits, deletions, reactions, drift) is an underappreciated input-surface problem for agents
- 3
Users can detect and reject model changes based on personality alone, even when the new model is cheaper and better on benchmarks
- 4
Shared integrations are a core architectural difference between personal and company agents; one connection should serve the whole team
- 5
Proactive agents trigger security backlash if deployed too broadly too fast; social acceptance requires graduated rollout
- 6
The Gmail incident reveals that customers mentally model AI agents as software, not employees, and product design must actively correct this
- 7
A 10-minute agent task feels slow in a web UI but fast in Slack because of expectation calibration, not actual speed
Key ideas
Slack is a superior interface for powerful agents because it reframes latency expectations
Why: Web apps train users to expect 30-second responses; Slack trains users to expect hours for human tasks, so a 10-minute agent response feels fast rather than slow
Implication: Agent interfaces should be chosen based on user expectation calibration, not just technical convenience
Company agents and personal agents are architecturally different in three dimensions: where they live, how they access integrations, and how memory is isolated
Why: Personal agents have one user, one memory, individual integrations; company agents have many users, shared integrations, and must prevent cross-channel context leakage
Implication: Building a company agent requires solving multi-tenant memory and permission isolation that personal agents never face
Model personality matters independently of task performance
Why: GPT-5.4 outperformed Opus on tool calling and code generation and was cheaper, but users rejected it because the personality felt wrong
Implication: Model selection for consumer/agent products must include subjective user experience metrics, not just benchmark scores
Proactivity must be earned, not enabled by default
Why: When Viktor proactively DMed users and joined threads on day one, security teams reacted negatively; gradual rollout with a small user base first prevents organizational rejection
Implication: Agent deployment strategies need social/political gatekeeping, not just technical capability checks
An AI employee should be treated as a hire, not a tool
Why: The Gmail incident occurred because the customer treated Viktor as software to be configured rather than as an employee who should not have personal email access
Implication: Product design and onboarding for company agents should mirror HR access-control workflows, not traditional SaaS permission models
Slack's interaction surface is non-linear and agents must linearize it
Why: Slack has DMs, channels, threads, edits, deletions, emoji reactions, and conversation drift between contexts; agents typically expect single-thread input
Implication: Building agents for rich communication platforms requires significant context-management infrastructure that chat-based agent architectures lack
Key facts
Viktor launched in February 2026
HIGHEvidence: We launched it in February this year
Viktor has no web app; it lives entirely in Slack
HIGHEvidence: lives in Slack, it doesn't have a web app
Viktor has access to 3,000 integrations and can build its own connections if needed
HIGHEvidence: it has access to 3,000 integrations and if for some reason it doesn't have access to your integrations, it can build its own connections
The company was founded in 2023 with the mission to build AI employees
HIGHEvidence: our mission from the very early days in 2023 was to build AI employees
Their earlier product, JCAI, was a browser-based web agent that took DOM snapshots and decided next steps
HIGHEvidence: it was taking a snapshot of your DOM minifying it in a lossless way and then based on the snapshot and this minified snapshot and your goal it was deciding on the next step
JCAI achieved 60% reliability for 3-5 step tasks in 2023
HIGHEvidence: it was working for like three to five five steps reliably. And by reliably I mean with 60% reliability
JCAI was state-of-the-art on the WebArena benchmark
HIGHEvidence: JCAI was a state-of-the-art web agent on on the most popular agenting benchmark called called web arena
Show 6 more facts
Their email agent product, Jace, is still active
HIGHEvidence: Jayce was like an amazing product... it's still alive
Viktor uses Opus 4.6 as its main model
HIGHEvidence: we use Opus 4.6 now for Victor
An A/B test switching from Opus to GPT-5.4 caused user backlash due to personality differences
HIGHEvidence: we wanted to try GPT 5.4 before... our users... they loved Opus and they all started raging when we did we did the AB test
One customer accidentally connected personal Gmail as a team integration, causing data leakage concerns
HIGHEvidence: One of the biggest e-commerce brands... they connected was their personal email, personal Gmail... this guy is texting me and saying, 'Hey man, like what the hell? Victor is leaking all of my data
Viktor now supports scoping integrations so they are not always shared
HIGHEvidence: we added a capability to Victor to kind of scope the integrations. So, they're not always shared
The speaker offers $100 in free credits to everyone in the room
HIGHEvidence: Everyone in this room has $100 in free credits
Quotes
“It is unworthy of excellent men to lose hours like slaves in the labor of calculation. Let us leave that to machines." — Gottfried Leibniz, quoted by Fryderyk Wiatrowski”
“Victor is not a tool. It's a hire.”
“if you hire a new employee, do you give them access to your personal email? Probably not, right?”
“our users... they loved Opus and they all started raging when we did we did the AB test”
“when you go to a web app and ask an agent to do something for you... you switched context and now you need to wait 10 minutes for the answer... It's quite frustrating”
“if you ping someone on Slack and tell them to build an app for you and get an answer in 10 minutes, you are shocked. No teammate has ever built you an app in 10 minutes, right?”
Verdict
This video is worth watching for anyone building or deploying enterprise AI agents. The unique signal is operational: it documents the specific failure modes that appear when scaling from personal to multi-user agents, with concrete examples (Gmail leakage, Opus personality backlash, security team resistance to proactivity). Most agent discussions focus on model capability or benchmark performance; this talk focuses on the organizational and UX friction that determines whether an agent survives in production. The Leibniz framing is lightweight; the value is in the war stories. COUNT: 13 facts, 0 assumptions, 0 demonstrations SIGNAL DENSITY: 78
Sources mentioned
"most popular agenting benchmark" — JCAI was state-of-the-art on it
Referenced as a personal agent example; contrasted with Viktor as a company agent
Mentioned as a viable integration connector for building your own agent
Example analytics tool Viktor can access to verify statistical significance of A/B tests
Example integration that only needs one person to connect for the whole team
Referenced as desktop-based alternatives; Viktor's advantage is cloud operation and shared context
Viktor's current main model
Tested as replacement; better on tool calling and code generation but rejected by users
17th century mathematician; quoted on automating calculation