Tech • AI • Robotics • Game
Building secure agents for knowledge work
5/10TL;DR
Every, a media and technology company focused on workplace AI adoption, has built a shared Slack-based AI coworker designed to spread best practices across organizations and turn expert judgment into reusable workflows.
KEY POINTS
From private prompting to shared workflows
Every argues that AI changes work less by replacing people outright than by forcing teams to redesign how they operate. In software, workers already use coordinated agents to accelerate roadmaps dramatically, and the company believes a similar shift is coming to broader knowledge work. The main barrier is not only model quality, but the difficulty of judging whether AI output is actually useful for a specific job.
A company agent inside Slack
The product, called Every Agent, is positioned as an AI coworker embedded in Slack rather than a separate personal assistant. The choice reflects a move away from one-agent-per-person setups toward a shared agent that multiple employees can improve together. By working in a common interface, staff can see prompts, responses and workflows in public, making AI habits easier to copy across the company.
Built for “AI ambassadors”
Every identified a recurring problem inside organizations: enthusiastic employees often discover powerful AI workflows, but struggle to persuade colleagues to adopt them. A shared Slack agent lets those users demonstrate results instead of merely describing them. That approach is meant to turn isolated experimentation into organization-wide practice.
Handling routine but painful work
One practical use case was open enrollment for employee benefits. Instead of having each employee upload a complex plan PDF to a chatbot individually, the company used the agent to understand the available plans, match them to each employee’s situation and guide plan selection. The value was not just time saved, but added confidence that recommendations reflected company context rather than a generic model response.
Encoding editorial judgment
A second example is “Kate Bench,” a system based on roughly 30,000 historical edits from editor-in-chief Kate Lee. Employees can ask the agent to do a “Kate pass” on a draft, and it will suggest edits directly in Google Docs in line with her preferences. Kate then reviews the changes, and the system tracks what she still had to fix so the editing skill can improve over time.
Automation as scale for experts
The company frames this not as replacing specialists, but as multiplying their reach. As the organization grew from 4 people to around 30, senior editorial review became a bottleneck. By converting expert taste into a reusable system, the company says one person’s standards can be applied across many more outputs without demanding more hours from that person.
Benchmarks tailored to real work
Every has become increasingly focused on custom evaluations rather than generic model leaderboards. Its internal process turns examples of good and bad model behavior into rule-based checks, closer to unit tests for knowledge work. Instead of asking whether a model scores well in general, the company asks whether it follows specific standards such as headline formatting, tone or editing conventions.
Why it chose managed agent infrastructure
Early attempts to run broader agent infrastructure in-house proved difficult. Every said Claude Managed Agents offered core features such as sandboxes, memory, session control and isolated environments, allowing the team to focus more on interaction design and workflow behavior than on infrastructure maintenance. The company also highlighted security benefits from built-in isolation and better control over what the agent can access on a user’s behalf.
The next step: compounding taste
The longer-term goal is a system that learns continuously from user corrections, preferences and examples. Every compares today’s frontier models to highly capable new graduates: impressive on tests, but unfamiliar with how a specific workplace operates. The aim is to build agents that gain “fidelity” to a user’s taste, judgment and working style over time, then make that expertise available to others in the organization.
CONCLUSION
Every’s approach reflects a broader shift in workplace AI from individual chatbots to shared systems that capture and distribute institutional know-how. The central question is no longer whether models can match human performance on general benchmarks, but whether they can reliably express the preferences and standards that define real work inside a company.
Ask a question