← All speakers

Bio, Work & Ideas

Fryderyk Wiatrowski

Conference affiliation: Viktor · 2026

Fryderyk Wiatrowski is the co-founder and chief executive of Viktor, an AI coworker for Slack and Microsoft Teams that uses a company’s existing software to complete everyday work. His defining ambition is a shared AI coworker that understands an organization without letting sensitive information travel indiscriminately between employees.

Wiatrowski studied mathematics and computer science at Oxford, led the Oxford University Computing Society, and organized hackathons and introductory programming initiatives. Before becoming a founder, he worked at Optiver and Meta.

In 2023, he and fellow Meta engineer Peter Albert founded Zeta Labs to build AI employees. Their early agent navigated websites by interpreting compressed representations of webpages, giving it access to software without dedicated integrations. But multistep browser automation proved slow and unreliable, prompting a shift toward Jace, an email agent that could schedule meetings, draft replies, and trigger actions such as refunds, with approvals where appropriate.

Viktor followed in February 2026, expanding this approach from individual inboxes to an entire organization. One employee can connect workplace tools that others access through a shared agent, subject to appropriate permissions. In May, Accel led Viktor’s $75 million Series A.

  • Automate operational noise, preserve human judgment. Hiring requires consequential decisions; scheduling interviews generally does not. Wiatrowski designs agents around reactive tasks triggered by emails, messages, or support requests, allowing people to retain oversight of higher-stakes choices.
  • Make organizational context useful without making it porous. A company-wide agent encounters executive discussions, engineering channels, private messages, and conflicting instructions. Memory isolation and access control determine what it can reveal or do. When a customer inadvertently shared a personal email integration across a team, Viktor added finer-grained integration scoping.
  • Design agents around workplace behavior. Slack threads, message edits, deletions, direct messages, and conversational tone affect how an agent should interpret tasks and retain context. Asynchronous messaging also changes expectations: work that feels slow in a standalone application can feel remarkably fast when delegated to a coworker. His AI Engineer presentation on Viktor connects those interface decisions with privacy and trust.

Read the topics behind these talks

2 conference talks

Key ideas

Scroll to read ↓

Viktor’s evolution from browser automation to a shared Slack agent shows why company context, conversation continuity, permissions, and personality matter as much as tool use.

  • An employee where the team already works
    0:15 ↗
  • From browser steps to an email-triggered agent
    2:08 ↗
  • Shared integrations create shared responsibility
    5:16 ↗
  • Slack changes what waiting feels like
    8:21 ↗
  • A Slack conversation is not a single thread
    10:07 ↗
  • Model selection includes personality
    11:55 ↗
  • Useful interruptions must earn permission
    12:55 ↗
  • Cloud execution and one company connection
    14:28 ↗
  • Personal Gmail exposes the scope problem
    15:44 ↗
  • What a coworker needs to become useful
    17:02 ↗
  • The invitation to try it
    18:44 ↗

Key ideas

Scroll to read ↓

Start with the work an agent can pick up without being asked, then build its capabilities through better prompts, explicit state, fine-tuning and feedback.

  • What would you hand to a reliable agent?
    0:33 ↗
  • Automate the work around the valuable decision
    2:27 ↗
  • Find the reactive layer
    4:45 ↗
  • From a trigger pool to browser actions
    7:28 ↗
  • Escalate only when the task requires it
    9:34 ↗
  • Make the prompt familiar and each decision smaller
    10:45 ↗
  • Track state explicitly and make updates easy to express
    14:02 ↗
  • Use structure without multiplying failure points
    16:06 ↗
  • Fine-tune on varied, demanding interactions
    17:08 ↗
  • Filter the data, then consider reinforcement learning
    18:30 ↗

References