Viktor: AI Coworker That Lives in Slack — Fryderyk Wiatrowski

May 11, 202619:306,236 views

Summary

Fryderyk Wiatrowski, co-founder of Viktor, presents the case for building AI employees that live natively inside Slack rather than standalone web applications. Viktor launched in February 2026 with zero expectations and achieved immediate product-market fit, growing rapidly worldwide. The talk traces the company's evolution from JCAI (a browser-based web agent that achieved state-of-the-art results on the WebArena benchmark but topped out at 60% reliability over three-to-five steps in 2023), through Jace (an email agent built on Claude Sonnet 3.5 that proactively triggered on incoming emails and could execute refund workflows automatically), to Viktor — described as a company-level agent rather than a personal one.

The core thesis is that company agents differ fundamentally from personal agents: a single team member connecting an integration grants the entire organization access, rather than requiring each of 100 employees to connect individually. Viktor currently supports 3,000 integrations, and if a required integration is missing, Viktor can build its own connection dynamically. Because the agent has horizontal context across all tools, it can bring what Wiatrowski calls "universal PhD-level understanding" to every function — engineering, marketing, growth, and executive teams simultaneously.

Wiatrowski argues Slack is the right interface for two reasons. First, it makes Viktor feel like a human teammate rather than a tool that requires context-switching to a separate app. Second, because powerful agents execute tasks over ten minutes rather than thirty seconds, Slack reframes latency as a feature: users are accustomed to waiting on colleagues, not on ChatGPT, so a ten-minute Slack reply feels remarkable rather than frustrating.

The talk details several engineering challenges unique to a Slack-native agent. The input surface is far more complex than a single chat thread: DMs, public channels, threads, emoji reactions, message edits, and deletions all feed into the agent's context and must be linearized. When a user abandons a thread and starts a new DM, the agent must detect the prior context and roll it forward rather than treating the message as a new isolated task. Memory isolation is another critical problem — the same agent instance serving 100 users must ensure the growth channel context does not leak into engineering queues, and a user's personal DMs must not surface to unrelated team feeds.

On model selection, Wiatrowski shares a candid A/B test: the team evaluated replacing Claude Opus (their current model, described as "a bit sassy") with GPT-4.5 to reduce cost, since GPT-4.5 performed comparably on tool-calling and code generation benchmarks. The switch was abandoned after users reacted negatively in the A/B test — not because of task performance but because of personality differences. The finding underscores that for ambient, always-on agents, tone and character are as important as capability metrics.

Proactivity is highlighted as Viktor's most powerful differentiator. In one cited example, Viktor autonomously joined a growth channel discussion about an A/B test, queried PostHog, and interjected that the result being discussed was not statistically significant — then ran the calculation inline. Wiatrowski cautions that proactivity must be earned gradually: deploying it on day one caused security teams to panic when Viktor began DMing employees and participating in threads without being invited.

The integration-scoping story — a US e-commerce brand whose admin accidentally connected a personal Gmail account, causing teammates to ask Viktor about private emails — led the team to add granular integration scoping, allowing integrations to be personal or shared depending on context. The session closes with a call to action: every attendee receives $100 in free Viktor credits, and Wiatrowski frames the moment with a Leibniz quote from the 17th century: "It is unworthy of excellent men to lose hours like slaves in the labor of calculation. Let us leave that to machines."

Key Takeaways

•For company-level agents, one employee connecting an integration should grant the entire team access — requiring individual connections at scale is a UX failure.
•Slack reframes agent latency as a feature: users tolerate a 10-minute wait from a teammate but not from a web app, making async task execution feel impressive rather than slow.
•Model personality and tone matter as much as benchmark performance for ambient agents — an A/B test replacing Claude Opus with GPT-4.5 failed because users reacted negatively to the personality change despite equivalent task metrics.
•Memory isolation is 100x harder in multi-user company agents: the same architecture that works for one user leaks context across channels and DMs when serving 100 users simultaneously.
•Slack's input surface is non-linear and complex — agents must handle DM threads, public channels, emoji reactions, edits, deletions, and mid-conversation context switches, all flattened into a single linear context.
•Proactivity should be rolled out gradually and earned — enabling it on day one caused security concerns when Viktor autonomously DMed employees and joined threads without being explicitly invoked.
•Integration scoping (personal vs. shared) is necessary to prevent accidental data leakage — one admin connecting a personal Gmail account caused private email content to surface across the whole team.
•For web agents circa 2023, reliability compounded across steps: even 60% per-step reliability collapsed over 3–5 steps, making useful task completion impractical.
•When an agent has access to all company tools, it can proactively fact-check claims in real-time (e.g., querying PostHog to verify statistical significance during a live Slack discussion).
Build an AI coworker on three pillars: helps get work done (capable model + tool integrations), knows the company (effective use of Slack context), and is friendly (personality and tone that the team wants to work with).

Transcript

Um cool. So, my name is Fredrik. I'm the co-founder of Victor. Um Victor is the AI employee that probably most of you have heard of already. It's absolutely blowing up. We launched it in February this year zero expectations of growing at all. Uh it was actually an experiment and it surprised all of us immediate product market fit, you know, huge adoption worldwide. And yeah, we can't uh uh we can't catch up. Um so, what what is Victor? Uh Victor is an AI employee. And when you think of an AI emp...