How Every is building secure AI coworkers for real knowledge work

Knowledge work is being rewritten in real time. Large language models are already strong on benchmarks, but the real question inside most companies is simpler and more practical: how do we turn these models into reliable coworkers that help real people ship better work, faster—and do it securely?

From benchmarks to real work

Modern models score impressively on standardized benchmarks, often matching or beating humans on pre-AI tasks. But benchmarks mostly answer the question, “Can this model do generic tasks?” not “Can it help me do my job better?”

Every’s view is that AI is most powerful when it extends a person’s capabilities rather than replaces them. Instead of obsessing over leaderboard scores, they focus on two questions:

• How do we use models to actually move our roadmap faster?
• How do we measure whether they’re helping on the work we really do?

Developers have been the first to feel this shift. With code assistants and agents, a single engineer can orchestrate a small “organization” of AI helpers and pull a roadmap forward by months or years. Every believes this pattern is coming to the rest of knowledge work—it’s just harder to verify success and to encode softer skills like taste, judgment, and style.

What Every actually does

Every is both a media and technology company focused on helping people work at the edge of AI. On the education side, they offer newsletters, courses, and in-depth “vibe checks” whenever a new model launches—hands-on, taste-driven evaluations of how a model performs on real workflows.

On the tools side, they’ve built the Every Agent: a company-wide AI coworker that lives in Slack and helps teams turn advanced AI workflows into shared, repeatable practices.

The Every Agent: an AI coworker inside Slack

The Every Agent is designed as an AI coworker that sits in your company Slack. Instead of each person tinkering with AI in private, the agent makes prompts, skills, and workflows visible and reusable across the organization.

Every built it with a specific persona in mind: the “AI ambassador.” This is the person who already has a finely tuned Claude or Codex setup that 10x’s their work, but can’t get the rest of the company to adopt it. A shared Slack agent lets them show, not tell—everyone can see great prompts and workflows in public channels and start using them themselves.

Key design ideas behind the Every Agent:

• It lives where people already work (Slack), not in a separate app.
• It acts like a shared coworker, not 20 separate bots no one else can see.
• It captures and spreads the best workflows so they don’t stay locked in one power user’s head.

Why Slack is the core form factor

Every originally experimented with a “one agent per person” model across multiple surfaces like Telegram and Slack. The problem: each person’s agent improved in isolation. Skills, instructions, and tricks didn’t spread across the team.

They realized agents get dramatically better when multiple people invest in the same one. Shared skills, shared memory, shared improvements. Slack was the natural home for that because it’s already the hub for conversation and collaboration.

Instead of copying and pasting between private AI chats and Slack threads, people can simply @-mention the Every Agent in-channel. The entire team sees how it’s used, learns from good prompts, and reuses working patterns.

Real-world use case: open enrollment, done by an agent

One of the most practical examples of the Every Agent in action is something nearly every company struggles with: benefits open enrollment.

Normally, employees get a dense PDF with multiple plans and have to do confusing math or call a more experienced friend or family member to understand what to choose. Yes, you could paste that PDF into a model yourself—but that still leaves every person doing the same work in isolation.

Every instead ran the entire open enrollment process through their company agent. The agent:

• Knew all the plan details.
• Knew who each employee was and their situation.
• Helped each person reason through which plan made the most sense for them.

This wasn’t just about saving time. Because the guidance came from a company-managed agent, employees trusted that the answers reflected both the raw plan data and the company’s perspective on what might be a good fit for different situations.

Automating an editor-in-chief’s taste

Every’s editorial team provided another powerful testbed: their editor-in-chief, Kate. She has extremely strong instincts for copy edits, structure, and style, and historically reviewed nearly everything before it went out the door.

As the company grew from four to thirty people, the volume of writing exploded—but Kate’s available time did not. Instead of simply working harder, Every turned her past work into an agent skill called “Kate Bench.”

They took roughly 30,000 of her historical edits and distilled them into a skill the Every Agent can run. Now, anyone can say “Do a Kate pass on this” in Slack, and the agent will:

• Apply Kate’s editing style to a Google Doc.
• Make suggestions in her voice and with her preferences (down to details like headline casing).
• Let Kate quickly review, accept, or reject edits rather than doing everything from scratch.

Every also tracks how much extra work Kate does on top of the agent’s edits and feeds that back into the skill. Over time, the “Kate Bench” skill compounds, catching more issues and better reflecting her evolving taste.

This flips the usual automation story. Instead of asking “Will AI take Kate’s job?”, the question becomes “How can Kate scale her impact without scaling her hours?” An agent that compounds her taste lets her influence far more writing while spending more time on the hardest, highest-value edits.

From vibe checks to personalized benchmarks

Every is known for its “vibe checks” on new models—hands-on evaluations of how a model performs on their real workflows. But doing this entirely by hand doesn’t scale.

To fix that, they started capturing real tasks and turning them into structured checks. Whenever someone likes or dislikes a model’s output, they log why. Over time, this builds a library of tests that look less like generic benchmarks and more like unit tests for their own work.

For example, a check for the Kate Bench skill might be “Is the headline in sentence case?” or “Does this email sound like our brand?” These checks let them run multiple models against the same real-world scenario and see which one best matches their taste and standards.

If you want to understand how this fits into the broader world of AI agents, it’s worth reading a deeper breakdown of what AI agent skills are and how they actually work.

Why they chose Claude Managed Agents

Before the Every Agent, the team tried building its own infrastructure for multi-agent systems. The lesson: infra is hard, and it quickly distracted from the real product work of designing great agent experiences.

With Claude Managed Agents, they get:

• Sandboxed tool use and code execution.
• Built-in memory and session control.
• A unified environment where the model “knows” how to use its tools.

This lets Every focus on interaction design—how the agent behaves, when it should speak up, how it should represent a person’s taste—rather than on spinning up sandboxes or maintaining brittle orchestration code. Upgrading to new Claude models also becomes a one-line change instead of a breaking migration across custom harnesses.

Security, isolation, and access control

Putting a powerful agent inside a company raises immediate security questions: what can it see, what can it do, and on whose behalf?

Every leans heavily on the isolation primitives in Claude Managed Agents. They can control:

• Which memories are attached to which sessions.
• What each sandbox can access.
• How sessions are created and scoped.

Conceptually, the Every Agent behaves like a coworker in Slack: a separate entity you talk to. Under the hood, when it reaches out to tools or company systems, it often acts “on your behalf,” respecting your identity and permissions. That allows nuanced policies—Dan and Willie may not see the same data, even if they ask the same question—because the agent understands who is asking and what they’re allowed to access.

If you’re designing something similar, it’s useful to think in terms of agents, tools, and permissions rather than just chatbots. A good primer on that angle is enterprise AI security explained: tools, agents, and access control.

Auth, identity, and “acting like a human”

Once an agent is inside your internal Slack, people naturally treat it like a colleague. They expect it to understand context, make reasonable judgment calls, and respect boundaries that are obvious to humans but tricky to encode as rules.

Every’s north star is to make the agent behave as much like a human coworker as possible within those constraints. That means:

• The agent is a distinct identity in Slack you can @-mention.
• It maintains isolated memories per user and per company context.
• When it uses tools, it typically does so as an extension of the requesting user, not as an all-seeing super-admin.

This design matches how people intuitively want to work with an AI coworker: clear internal vs external boundaries, but flexible enough that the agent can still be genuinely useful.

Continuous improvement through everyday nudges

One of the most interesting aspects of the Every Agent is how it learns from tiny interactions. Not every improvement comes from a formal eval—many come from casual corrections like “That joke wasn’t great, try another” or “This tone is too formal.”

Those micro-adjustments become training data for skills and preferences. Over thousands of examples, they distill into a representation of a person’s taste that’s much richer than any single prompt could capture.

The pattern looks like this:

• Capture real work (drafts, edits, responses).
• Capture feedback (what you liked or disliked and why).
• Turn that into checks and skills.
• Let the agent apply those skills across the org.

Instead of manually doing the same task forever, experts gradually shift to working on the system that does the task for them.

Agents as new hires: from SAT scores to work trials

Every uses a helpful analogy: hiring an AI agent is like hiring a new grad. Benchmark scores (like SATs) tell you they’re smart, but they don’t tell you how they’ll perform at your specific job.

On “day one,” the model is a very bright new hire who has read a lot but doesn’t know how you work. Over time, as it gains experience on your tasks and your feedback, it develops “fidelity” to your preferences and standards.

That fidelity—how closely the agent’s work reflects what you actually want—is what unlocks the next leap in productivity. The goal is to reach a point where you’re not doing every unit of work yourself; you’re managing the system, intervening like a manager when needed, and letting the agent handle the rest in your style.

What they want next from models like Claude

As models get more capable, there’s a risk that raw intelligence outpaces what everyday knowledge workers can actually absorb. Every’s wishlist for model evolution is less about exotic capabilities and more about usability:

• Responses calibrated to the user’s level of knowledge.
• Outputs that reference things the user will actually remember or understand.
• Stronger support for consistent, configurable personality and tone.

They also care a lot about personality design. Different customers want different vibes from their AI coworkers—some prefer a more aggressive, proactive agent; others want something patient and understated. Being able to dial that in a structured way (rather than via brittle prompt hacks) is high on their list.

Where work is heading in the next year

Looking ahead a few months to a year, Every expects more knowledge workers to adopt the “loop” mindset developers already use with code. Instead of manually doing every step, you orchestrate loops of agents that execute and refine work, while you manage quality and direction.

To make that viable at scale, agents need:

• High fidelity to individual and organizational taste.
• Strong security and identity boundaries.
• Interfaces that fit naturally into existing workflows (like Slack, email, docs).

As those pieces mature, experts inside organizations will be able to spread their skills and taste far more widely without multiplying their hours. Colleagues and clients will be able to “work like them” by using skills encoded in shared agents.

The unexpected effect: more work, not less

Despite fears that agents will eliminate work, Every’s internal experience has been almost the opposite. The more they use agents, the more new kinds of work appear—higher-quality projects, more ambitious ideas, and deeper polish that would have been impossible under old time constraints.

The nature of the work changes, of course. People move from doing every task by hand to designing, supervising, and improving systems that do those tasks. But instead of a barren horizon where there’s nothing left to do, they see a new horizon full of higher-level work that was previously out of reach.

The small magic moments

For all the talk about infra, evals, and security, some of the most convincing proof that agents are becoming real coworkers comes from small, human-feeling moments.

At Every, the agent tracks “beer debts” their CEO owes people and surfaces them at just the right time in conversation. Another skill mimics a teammate’s distinctive all-caps, over-the-top announcement style so others can send messages in his voice. These touches aren’t just fun—they signal that the agent understands enough context and personality to feel like part of the team.

That, ultimately, is the bar for AI coworkers: not just intelligence, but familiarity, taste, and trust. The Every Agent is one of the more concrete attempts so far to hit that bar inside real organizations—and a glimpse of how knowledge work may look once agents are part of every team’s daily rhythm.