Was this newsletter forwarded to you? Sign up to get it in your inbox.

Signal

There’s more than one way to hire an agent

Shopify has River to help engineers ship code. Stripe has Kai to turn company data into dashboards and documents. We’re building Every Agent to encode our team knowledge into a shared agent. It’s official: The age of the company-wide agent has arrived.

The phrase makes this sound like a new software category. It is really a spectrum of ownership. A company can build the whole system, rent the machinery underneath it, or buy an agent that already lives in Slack or Notion. The right choice for your organization depends on what you want it to do, where you want it to live—and how much of the upkeep you’re willing to take on after launch.

Putting a bot in Slack is the easy part—at least that’s what we’ve found building Every Agent. Most of the work of designing the agent sits behind the interface: deciding which sources of data are authoritative, maintaining the connections, teaching the agent how the company works, and limiting what it can do without approval. That is why products that look similar in a demo can require very different commitments from the customer.

Shopify and Stripe sit at the high-ownership end. River draws on Shopify’s single version-controlled repository, consistent development environments, public Slack, and a process for identifying repeatable workflows that can be turned into skills. Stripe built Kai on LangChain, a platform that makes open-source tools and managed infrastructure for building and running AI agents. LangChain says one engineer shipped the first version of Kai in a week, but that speed rested on more than a decade of the company’s internal tooling and security infrastructure. Kai now has more than 500 tools and 1,000 skills. The tools connect Kai to Stripe’s data warehouse, intelligence dashboards, and project-management systems, while the skills cover jobs such as researching an account ahead of a sales call or triaging a billing escalation. Build at this level when the agent’s advantage comes from proprietary systems—and when your company can keep its tools, skills, and permissions working.

LangChain serves two parts of this market. Stripe used Deep Agents, the company’s open-source framework, to build a system it largely owns. Managed Deep Agents is for companies that still want a custom agent but don’t want to run its infrastructure. LangChain runs persistence, memory, skill loading, sandboxes, deployment, and evals. The company using it still supplies the models, prompts, tools, and rules.

Notion, Lindy, and Viktor sit farther toward the buy end of the build-to-buy spectrum. The vendor takes care of more of the agent infrastructure while the customer is responsible for the knowledge, instructions, and permissions that agent should have.

Notion Custom Agents make sense when the relevant knowledge already lives in Notion and the job has clear inputs and outputs: Prepare weekly priorities, triage feedback, update a database. Lindy offers more control over a known Slack workflow, including its triggers, channels, filters, and knowledge base. Viktor goes broad, promising one shared agent across Slack or Microsoft Teams with connections to thousands of tools. The more of the system a vendor supplies, the faster a team can start—and the more carefully it should examine the vendor’s memory, permissions, and output quality.

Companies will likely mix these approaches. Kai already does: It is a custom agent built on a vendor framework. A company can expose one shared agent to employees while running specialist agents and managed infrastructure behind it. “Company-wide” may describe the front door more often than the system behind it.

What to do: Before you compare vendors, write down four things: the platforms or workspaces where colleagues will use the agent (for example, Slack, Notion, or Teams), which company knowledge it needs, who will maintain that context, and which actions require approval. Those answers will tell you what to build, what to rent, and what to buy off the shelf.

Steal this workflow

Keep your coding-agent sessions alive

Closing a terminal window shouldn’t kill a long-running agent job—or your train of thought. Every head of video Randy Counsman tried Herdr after accidentally quitting Warp a few times and reopening it to a mess of color-coded tabs, split windows, and dead sessions. Herdr keeps sessions running when he closes the client, asks him to name tabs as he creates them, and lists his agents in the bottom-left corner, making it easier to pick up where he left off.

You need: macOS or Linux, Homebrew, a terminal, and a project where you already use Claude Code or Codex. Native Windows support is still in preview. Start with a non-sensitive repository.

Install Herdr.

brew install herdr

Start a Herdr session from your project.

cd /path/to/project

herdr

Run your coding agent inside the session.

claude

Run codex instead if that is your agent.

Detach and come back. Press ctrl+b q, close the terminal window, and later run herdr from the project again.

Run herdr server stop when you mean to end the live session. The catch: Closing the terminal window is safe; stopping Herdr is not. If you detach, your programs keep running. If you stop or restart the server, only the window layout returns. Pane history is off by default because terminal output can contain passwords, tokens, prompts, and command output.

Try it: On a low-stakes repository, ask the agent to run the test suite and summarize failures. Detach while it works, then reattach. You will learn exactly what survives before relying on it for anything important.

Discuss

“More time should be spent writing a document than consuming it.”

—Clay’s company-wide AI writing policy, written by Sophie Alpert for its engineering team and shared by cofounder and head of operations Varun Anand after Clay expanded the policy company-wide.

Clay’s rule acknowledges the reality of what a world inundated with AI-generated text feels like. AI now lets an author generate a document much faster than a colleague can read it. The policy does not ban AI. It makes the author responsible for the thinking, the editing, and every sentence they circulate.

We at Every have published our own editorial guidelines for writing about and with AI. Like Clay’s, our guidelines don’t preclude the use of AI, but they stress that a human writer must stand behind every word of their content, no matter how that content was produced.

It takes time to establish norms around new technology. With stories like Clay’s, we’re seeing how those norms get shaped one policy at a time.

What we’re reading

Zuckerberg’s plan for personal superintelligence

Mark Zuckerberg’s new statement spells out Meta’s AI policy: Put a personal agent in the hands of billions of people, largely through products Meta already owns.

“The Future is [sic] for Everyone” says superintelligence should work for individuals. In Zuckerberg’s version of the future, everyone gets a personal agent that knows what they care about and helps with work, money, health, relationships, and creative projects. Spreading systems that powerful across billions of people, he argues, is safer than letting a few labs, companies, or governments control them. He makes a similar case about jobs: The transition goes better if AI expands what people can invent and build before it automates their existing work.

Meta plans to give billions of people and small businesses access to personal superintelligence, keep it free or cheap, resume some open-source releases, and create a private mode that Meta itself cannot inspect. Zuckerberg says Meta’s independent board will approve the safety criteria for model releases and review whether each release meets them. He also wants frontier labs to give the U.S. government early access to model checkpoints so it can prepare for security risks without delaying public releases.

Several of those promises match what Meta is already doing. Meta AI can connect to email and calendars, make plans, and act on a user’s behalf. Muse Spark, Meta’s large language model, is moving into WhatsApp, Instagram, Facebook, Messenger,

and Meta’s glasses. The company expects to spend $115–135 billion on capital expenditures this year, with much of the increase going toward its superintelligence lab and core business. The open-source promise is murkier. Muse Spark is still limited to selected API partners, and Meta says only that it hopes to open-source future versions.

Zuckerberg calls this plan a distribution of power, yet Meta would still own the apps, models, and infrastructure through which much of that power arrives. He does not address that tension in his statement.

Links worth a click

In Anthropic’s controlled test of Claude Code auto mode, paid professional testers caught a substituted dangerous command only 13.6 percent of the time. METR’s blueprint for independent investigations lists what an outside researcher would need to assess whether a company’s agent lies, cheats, or slips its safeguards. OpenAI’s academic-researchers announcement puts numbers to the kinds of work scientists are handing to AI: Heavy users were almost twice as likely to submit tasks estimated to take four hours or more. The AI Agent Index compares 30 deployed agents across 45 fields, including autonomy, safety, architecture, and transparency. Browse it when you have half an hour to lose responsibly.

Katie Parrott is a staff writer at Every. You can read more of her work in her newsletter. To read more essays like this, subscribe to Every, and follow us on X at @every and on LinkedIn.

Everyone’s a builder now. Every All Access gets you the full membership plus the Builder Pack—$9,000+ in credits for the tools we build with.