Source: raw/Building_secure_agents_for_knowledge_work.md — official Anthropic YouTube channel (@claude), “Building secure agents for knowledge work”, a conversation with Dan (Shipper, Every’s CEO) and Willie of Every, youtube.com/watch?v=z7cNbsr3b5s (fetched 2026-10-30). The Anthropic interviewer is not named, and the transcript has no speaker labels, so most claims are attributed to “Every” below.

Every, a media and technology company, built the Every Agent: one AI coworker per company, living in Slack and running on Claude Managed Agents. The talk covers why Every moved from per-person agents to one shared agent, how it splits identity (a separate entity in conversation, acting on the user’s behalf for tool calls), which Managed Agents primitives it relies on for isolation, and how it turns an expert’s edit history into a skill and a personal benchmark (“Kate Bench”). The talk is about adoption and design more than security mechanics. The security content is mainly about isolation and identity.


Key Takeaways

Why one shared agent in Slack

  • Started with OpenClaw on Mac minis, then consolidated. Every began “in the OpenClaw craze” with one agent per person on whatever surface they liked (Telegram, Slack). Those agents “died off because they’re just really hard to maintain.”
  • Shared investment compounds; individual investment doesn’t. “Agents got better… the more you invested in them,” but one person’s gains “weren’t really spreading out to the rest of the team.” Real work happened “when there were multiple people investing in them, making better skills, giving it better guidance.”
  • Prompting in public spreads practice. The “AI ambassador” (the colleague whose own setup is 10x-ing their work but whose coworkers tune out) can “show instead of tell,” because in Slack “you’re prompting in public instead of in private.” It also removes the copy-paste between a thread and someone’s private Claude Cowork session.
  • Small, coordinated tasks are the everyday win. Every ran open enrollment through the agent: it knew all the benefit plans and who each employee was, and helped each person pick. Leaders can write skills once and push them to the whole org.

Building on Managed Agents

  • “We don’t want to be an infra team.” Their first one-agent-for-everyone version “got unwieldy very quickly” while they were also trying to improve the product. One founder admits he thought infra was “not that hard anymore. We’ll just have Claude spin up a bunch of servers”; a month later: “This is hard.” Managed Agents “gave us sandboxes, it gave us memory, it gave us session control.”
  • Clone production state into a staging agent. They run one managed agent as an “edge” or staging instance, copy over “the memories and the durable state that we have from production”, and run experiments in isolation on the same infra (conversational gray-area cases, cutting token cost for customers).
  • Model upgrades are a one-line change. On the day of filming a new model launched with breaking changes in its prompting guide; “in Claude Managed Agents, it’s just a 1-line update.” Their view: “there’s not a clear separation between the model and the harness anymore. A model that doesn’t come with a computer that it knows how to use, it’s just not as performant.”

Security: isolation and identity

  • Isolation primitives give per-request control. Built-in isolation of “the sandbox, the tool calls… what are the environments, and where are the memories” lets them decide “who accesses what, when.” The test case they give is: “if Dan requests something, does he get the same thing that I would get?” They answer it by choosing which memories and which sandbox attach to each request.
  • Two identity modes. The North Star is to “act as close to a human as it can within Slack.” In conversation the agent “acts as a separate entity” with “isolated memories and an understanding of who you are and what you have access to.” When it fetches something through a tool call or connection, “it, for the most part, acts on your behalf… as an extension of you.” Every says users accept this split between internal and external.
  • The gray zones are hard to encode. People treat the agent like a human colleague and ask for things where “a human would know that there’s no harm in me being able to access this report”, but “that’s hard to encode into a piece of agent logic.”

Taste capture and personal benchmarks

  • Kate pass. Dan took “30,000 of her historical edits” from editor-in-chief Kate and turned them into a skill. Anyone can ask for “a Kate pass” on a piece, landing page or email, and the agent leaves suggested edits in the Google Doc as her. Kate reviews them, the extra edits she adds are tracked and “compound[ed] back into the skill”, and she gets a graph of time saved plus suggested new rules. Context: Every has grown from 4 people when Kate joined to 30, and publishes “10 times the amount of things.”
  • Experts scale without spending more time. This changes the automation question from “is it going to take away my job?” to letting an expert “spread your impact… without having to spend more time.”
  • Turn nudges into checks. Every captures the moments a user liked or disliked a model’s output and talks with an internal platform about why. The platform turns that into “checks, which are basically unit tests or rules” (for example “is the headline sentence case?”), which can then run “many different models on the same scenario.” This adds a quantified layer to their manual vibe checks. The platform “will eventually be part of the Every Agent.”
  • SAT score vs work trial. Benchmarks are like SAT scores: useful when scores differ, but “if everyone got a 1600, which is where we are with these models,” you need “a reference check… a work trial.” The goal is an agent that is “better on day 10 than on day one.”
  • Small corrections are the raw material for taste. “It’s hard to describe your taste all in one big go, but it’s easy to take it from a thousand examples and distill it out.” Every says this applies to editing, slide decks and email.

Product direction and asks

  • Hands. An experiment in which the Every Agent drives a computer-use agent on a team member’s desktop (“@Every, use Hands to do it on my computer”). Every says it fits alongside Claude Code and Cowork: public work in Slack, private work handed off seamlessly.
  • Social norms and personality are the open problems. Claude has been trained to answer every message it sees; in a channel with several humans (and other Claudes) it needs to know when to interject, when to connect two threads, and whether to be “an extrovert or an introvert.” Every wants personality to be “consistent but also configurable” per customer, and joked about an API parameter for “personality aggression.”
  • “Fidelity” is their word for the target. The agent should have “fidelity for what you want, what you care about”, so that, as in coding, you work “on the loop” and intervene the way a manager would.
  • More agent use has meant more work. “The more we use agents, the more work there is to do.” Their ask of Anthropic: don’t forget knowledge workers while frontier gains go to math and biology. Responses should be understandable and pitched at the reader’s level.

Try It

  1. Consolidate personal agents into one shared agent in your team chat. Use a Slack-resident agent such as Claude Tag so that skills and corrections build up in one place.
  2. Build an expert-pass skill from history. Export one expert’s past edits or review comments (Every used 30,000), turn them into a skill, have it make suggested edits rather than final ones, and log what the expert still changes so you can feed it back.
  3. Turn your next ten likes and dislikes into checks. Each time a model output pleases or annoys you, write one checkable rule (“headline in sentence case”). Run two models against the same scenarios and score them on your rules instead of a public benchmark.
  4. Decide your identity split explicitly. Write down when the agent speaks as itself and when it acts with the requester’s permissions. Then test the “does Dan get what I get?” case for one sensitive document.
  5. Keep a staging agent. Before changing prompts or models in production, copy production memory into an isolated agent and run the change there first.


Open Questions

  • Willie’s surname and role are not stated. Many lines can’t be tied to a speaker because the transcript has no labels.
  • No figures are given for the Kate-pass time savings in this talk. The earlier “12% less work” number comes from another talk. “30,000 edits” (this talk) and “three years of edits” (earlier talks) are compatible but describe the dataset differently.
  • The “new model launch” on filming day and its breaking prompt changes are not named.
  • How the on-behalf-of identity is implemented (OAuth per user, vault credentials, or something else) is not described.
  • Whether the Every Agent uses Managed Agents’ own memory stores or Every’s own memory layer is not fully clear. They say Managed Agents “gave us memory” and talk about copying “memories and durable state” between instances.