How Every builds a shared AI coworker that learns a team’s style

Katelyn Lesse asks how the tools people need are changing as AI changes work. Every’s Dan Shipper says the team tries the models on problems it actually has rather than choosing between fear and an imagined work-free future. In practice, he says, adopting the tools means reinventing a workflow through use.

Generic benchmark progress can look like a measure of whether AI can replace a person. Shipper’s view is that, used well, the tools extend a person’s influence and abilities. He wants ways both to measure that extension and to help people learn how to achieve it.

Programming has moved first because it is easier to check whether code works. Shipper describes engineers coordinating teams of agents and moving work forward on their roadmaps. He expects similar changes in other knowledge work, but says progress has been slower because quality is harder to verify and the working style is less familiar.

Every is a media and technology company whose offerings include a newsletter, courses and hands-on assessments of new models. Its product, Every Agent, is an AI coworker intended to bring the team’s way of working to other companies.

Shipper says the team initially adopted individual OpenClaw agents, but they were difficult to maintain and eventually disappeared. A single company agent proved more useful: workflows discovered by one person could spread to coworkers. Even a company focused on new AI tools found it hard to keep everyone working at the same level.

He describes the “AI ambassador”: someone who has built a valuable personal setup but struggles to persuade coworkers to try it. In Slack, people can show the agent what to do in public instead of explaining an invisible private workflow. The agent can handle recurring jobs such as gathering reports or preparing social posts, while colleagues see how those requests work.

Asked for a concrete example, Shipper chooses a routine task: benefits enrollment. Employees normally have to work through a PDF of plans and calculate which option suits them. He says the company agent knew the plans and the individual employee’s situation, and helped people make their selections. A shared agent also let company leaders distribute the relevant skill to everyone.

Lesse notes that a shared process saves each person from separately pasting the same document into a chatbot. Shipper says having the company prepare guidance for each person also made him more confident in the choice; navigating the options had been complicated enough that he used to ask his father for help.

Willie Williams describes another recurring workflow in the editorial team: a tool called Kate Bench, named after Every’s editor-in-chief, helps with the top edit before publication. Instead of filling Kate’s inbox, colleagues can tag the agent in Slack for a review. Kate can see the result and judge whether it is doing the job.

Shipper says Kate joined when Every had four people; the company now has about 30 and publishes roughly ten times as much. She still needs to ensure that copy edits meet her standards. To spread that work, he used 30,000 of her historical edits to create a skill in Every Agent.

Anyone can request a “Kate pass” on an article, landing page or email. The agent proposes changes in Google Docs, which people can accept or reject. Kate then reviews the result; Shipper says it catches most of what she would change and sometimes things she might miss, leaving her less work per piece.

The team tracks what Kate adds after the agent’s pass and feeds that back into the skill. She can see an estimate of the time saved and potential new rules. Shipper argues that the aim is to let an expert spread the way they work without having to spend proportionally more time on every request.

Lesse asks how the team measures and improves work like Kate Bench. Shipper says the important question is not a model’s score on a generic benchmark but how it performs on a particular person’s job. Every’s manual “vibe checks” try new models on its own workflows and report the team’s preferences, but those tests take time.

The team is capturing examples encountered during daily work—outputs people liked or disliked—and discussing why. Its platform turns those judgments into checks, such as whether a headline uses sentence case. Running different models on the same scenario then lets the team compare performance according to the standards that matter in its real work.

Every originally tried giving each person an agent on surfaces such as Telegram and Slack. Williams says the team moved toward one shared agent because an individual’s improvements were not spreading. The strongest results came when several people invested in better skills and guidance for the same agent.

Slack was a natural place for the shared agent because that is where coworkers were already talking. With separate agents, a collaborative discussion could require people to copy messages into a private tool and paste the answers back. A shared conversation keeps the people and agent in one place.

Lesse asks why Every built on Claude Managed Agents. Williams says the first attempt taught the team that infrastructure is hard: improving a product while maintaining a durable system to run agents quickly became unwieldy. Every wanted to focus on the future of work rather than become an infrastructure team.

The managed service supplies isolated computing environments, memory and session controls, Williams says, leaving more time for agent interaction design. Shipper recalls initially thinking the team could simply have Claude run the servers; after a month of trying, he agreed with Williams that the infrastructure work was difficult.

Williams says having those building blocks together also makes experiments easier. The team can copy production memories and persistent state into a separate test environment, try edge cases and explore how to reduce token costs without affecting the main product. The test runs on the same underlying infrastructure as production.

Many of those experiments concern judgment in conversation. People assume a coworker will make reasonable calls in ambiguous situations, but it is hard to express those expectations as instructions. Every combines evaluations with a more subjective question: does this interaction feel like working with a real colleague?

On security, Williams points to built-in isolation. The team can separate the sandbox, tool calls, environments and memories, and use session management to decide who can access what. Those controls let it consider whether a request from Shipper should receive the same information as an equivalent request from someone else.

Shipper adds that he sees the model and the surrounding system as increasingly intertwined. In his view, a model works better when it comes with a computer it knows how to use. He says a newly released model required changes to prompting, but moving to it in the managed service was a one-line update; Lesse says the service incorporates those adjustments.

Lesse asks whether the agent acts as a person, on a user’s behalf or on its own behalf. Williams says internal company requests often fall into gray areas where a human might infer that access is harmless, but that judgment is difficult to encode. The team’s goal is for the interaction to feel as close to a human coworker as practical.

In Slack, the agent acts as a separate entity in conversation. Behind that interaction it has isolated memories and an understanding of the requester and their company access. When it retrieves a document or looks something up through a connected tool, it generally acts on the requesting person’s behalf. Williams says users have found that boundary intuitive.

For improvement over time, Williams looks at the small corrections people give during work: a sentence that was not written well, or even a joke that did not land. A person’s taste can be hard to state all at once, he says, but many examples can reveal it. Those examples can become the basis of skills for editing, presentations or email.

Shipper compares evaluating an agent to hiring. A standardized test score can be a useful initial signal, but if candidates all score highly, an employer still wants references and a work trial. An agent on its first day may be well trained and capable but does not yet know how a particular company works.

Every wants to collect examples of good and bad work, draw out a user’s preferences and measure whether the agent improves with experience. Shipper says that ability will not be in the agent on day one; it is where the team hopes to go. Instead of repeating each task manually, a person could work on the system and see how it improves.

Asked about the next product ideas, Shipper describes personal benchmarks: a way to record what someone considers a good response, compare models on that basis and improve skills for their work. He also wants those skills to spread through an organization so other people can benefit from an expert’s approach without requiring more of that expert’s time.

Williams describes an experiment called Hands, a computer-use agent on his desktop that Every Agent can ask to operate the computer. The team sees that as complementary to personal tools such as Claude Code or Cowork: sometimes people want to collaborate in public, and sometimes they want to move the work into a private session.

Managing context becomes different when an agent joins coworkers. Williams asks whether it was present for a conversation, whether it should bring together ideas from separate threads and whether it understands when an intervention is appropriate. He sees interaction and personality design as a major part of making a group agent useful.

Shipper notes that a one-to-one assistant is usually expected to answer every prompt, while a social setting may include several people and other agents. It needs norms for when to speak or stay quiet. Every also wants the product to reflect its own way of working, so that using the agent feels like working with someone from the company.

Looking ahead, Shipper expects knowledge workers to manage loops of agent work much as some developers already coordinate coding agents. He calls the desired property “fidelity”: an agent should reflect what a person wants and values without requiring their involvement in every step. Generic benchmarks do not measure that personal fit well.

In that future, he says, individuals would work on the system and intervene at selected moments as a manager might, rather than performing every unit of work. Other colleagues or clients could then use an expert’s approach without taking up more of that expert’s time. He presents this as a direction for the coming year, not a completed capability.

Asked what Every needs from Anthropic, Shipper says greater model intelligence is valuable but knowledge workers also need answers they can understand, pitched at the right level of knowledge. Williams adds speed and cost, along with a way to keep an agent’s personality consistent while allowing customers to configure it—for example, to be more assertive or patient.

Shipper’s assessment of Every’s experience is that the more the company uses agents, the more work it finds to do. He expects the effects to differ across companies and parts of the economy, but says automation has shifted where people work in the system rather than necessarily making the work disappear.

For a lighter example, Williams says Every’s CEO asks the agent to remember when he owes someone a beer; the agent occasionally brings the debt up in a relevant conversation. Shipper says a colleague named Andrey writes funny, emphatic company announcements, so he made a skill that lets the agent draft announcements in that style. Coworkers now use it for their own announcements.