Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Viktor: AI Coworker That Lives in Slack — Fryderyk Wiatrowski
TL;DR
Viktor hit immediate product-market fit after launching in February 2025 — Fryderyk Wiatrowski says the Slack-native “AI employee” started as an experiment with “zero expectations” and then saw huge global adoption they “can’t catch up” with.
The core bet is that a company agent should live in Slack, not a separate app — Viktor joins channels, threads, and DMs like a teammate, uses 3,000 integrations, and can even build missing connections so it works where teams already coordinate.
Wiatrowski’s team learned the hard way that browser agents were too unreliable to be useful in 2023 — their earlier product Jace AI could only complete roughly 3–5 steps at about 60% reliability on WebArena-style tasks, which made failures and latency compound fast.
Turning a personal agent into a company-wide coworker creates nasty memory and permission problems — Viktor has to keep context from executive, growth, engineering, and private DMs separate, while still remembering ongoing work across messy Slack behaviors like edits, deletions, and abandoned threads.
Model personality mattered enough to overturn a cost/performance win — even though GPT-5.4 looked strong on tool calling and code generation, users “started raging” in A/B tests when Viktor switched off Opus 4.6 because they preferred Opus’s tone and slightly sassy personality.
Viktor’s strongest differentiator is shared company context, but that also changes the security model — one integration connection can serve an entire team, which is powerful for things like Meta Ads or PostHog, but it also led one ecommerce customer to accidentally expose a founder’s personal Gmail until Viktor added scoped integrations.
The Breakdown
From side experiment to “we can’t catch up”
Fryderyk Wiatrowski opens with the headline: Viktor launched in February with basically no growth expectations and immediately took off. His framing is simple and memorable — this isn’t an AI tool, it’s an “AI employee” that lives alongside the team in Slack and already has broad company context.
The long road from browser agents to something usable
The original 2023 vision was to build AI employees through browsers, back when tool calling and strong codegen models didn’t really exist. Their first product, Jace AI, used DOM snapshots and minified page state to decide actions, but in practice it could only do around 3–5 steps at about 60% reliability, which made it impressive on benchmarks like WebArena and still painful as a product because you’d wait a minute just to watch it fail.
Jace the email agent was the bridge to Viktor
After that, the company shifted Jace into an email agent that triggered when new mail arrived, connected to tools, drafted responses, and even took actions like issuing refunds with optional approvals. Wiatrowski describes this as the first real step toward agents that don’t wait in a web app for instructions, but instead already have enough context to proactively help.
Why a company agent is different from a personal agent
He draws a sharp line between personal agents like OpenClaw and company agents like Viktor: with Viktor, one person can connect an integration and the whole team can use it under the right permissions. That unlocks shared context across 3,000 tools, but also introduces hard problems around memory, conflicting instructions, and making sure sensitive context from an executive or growth channel never leaks into engineering, support, or the wrong DM.
Slack is the right interface — and a chaotic one
Wiatrowski says Slack fits because human coworkers don’t live in web apps, and because long-running tasks feel normal there — if someone builds you an app in 10 minutes on Slack, that feels fast, while waiting 10 minutes in a dedicated agent UI feels broken. But Slack also explodes the complexity: DMs, public channels, threads, emoji reactions, edits, and deletions all become agent inputs, and Viktor has to turn that chaos into something like a coherent working memory.
The weirdly important thing: tone
One of the best stories in the talk is that users revolted during an A/B test when the team tried replacing Opus 4.6 with GPT-5.4, even though GPT looked great on tool use, codegen, and price. The issue was personality — users loved Opus, found Viktor funnier and more human with it, and Wiatrowski half-jokes that Viktor’s slight sassiness may be part of the product.
Proactivity is powerful, but only after trust is earned
Viktor can jump into a growth conversation, check PostHog, and say a claimed experiment result isn’t statistically significant — then explain the math. That’s the dream, but he says rolling this out too early backfires: if Viktor starts DMing everyone and chiming into threads on day one, security teams panic, so proactivity has to be earned with a small trusted group first.
“It’s not a tool, it’s a hire”
His closing story lands the product philosophy: a large US ecommerce customer connected a personal Gmail as the first team integration, then got angry when coworkers started asking Viktor about those emails. Wiatrowski’s response was basically: if you hired a new employee, would you hand them your personal inbox? That mistake led them to build scoped integrations, and he ends by summarizing the recipe for a great AI coworker: help get work done, know the company, and be friendly enough that the team actually likes working with it.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
