Rails

ThinkingNovember 2026

Agent management is the new IT department

Companies that gave every desk a computer ended up with an IT department. Agents need one too, and most companies lack it.

Every company that put a computer on each desk ended up with an IT department. The machines needed buying, connecting, securing and upgrading, and somebody had to own all of that.

Agents are reaching the same point now. Companies want them doing real work, and the models are smart enough to do it. What nobody owns yet is the job of keeping them working.

Then: every desk got a computerIT showed up to run themITNow: every person gets agentsNobody runs them yet

Running agents is a job of its own

An agent has four parts: the playbook it follows, its memory, the tools it can use, and its own app, the pages it pulls up beside the conversation for the work at hand. The chat it runs in is the host. Each of those parts needs someone to set it up, keep it current and decide who gets to touch it.

Today that work is split across the company. Token spend sits with finance, agent identity with security, and the protocols agents use to talk to each other with engineering. Nobody holds the whole picture: which agents exist, what each one may do, who it works for and what it costs.

A tool can fill one slot in that job, like a dashboard for spend or a vault for keys. The rest needs a person who answers for all of it, the way an IT department answered for every machine in the building.

Every old IT job has an agent version

Picking vendors is the clearest case. IT chose the hardware and the software a company ran on, and someone now has to choose the models, tools and open source pieces each agent runs on.

Access works the same way. IT issued accounts and decided who could open which system. Each agent now needs its own identity, a set of tools it has been granted, keys it never sees and a spend cap that pauses it at the limit.

The help desk and the overnight upgrade carry over as well. When an agent needs a person, its question should reach whoever owns that work, in one place. When a better model ships, someone swaps it in underneath the agent while the work keeps going.

IT jobAgent version
  • Pick vendors and buy hardwarePick the models, tools and open source pieces each agent runs on
  • Issue accounts and passwordsGive each agent its own identity, granted tools and keys it never sees
  • Keep the license budgetA spend cap per agent that pauses it at the limit
  • Help deskA question that needs a person goes to whoever owns that work, in one place
  • Overnight upgradesSwap the model underneath while the agent keeps working
  • Asset inventoryOne list of every agent, what it may do, who it works for and what it costs

I run this department for my own agents

I learned this by doing it. My recruiting and the product itself run through my own agents, and over the last few weeks I have set the rules an IT department would set.

The volume explains why those rules matter. Most days I work with my agents in two shifts, a long one through the working day and a second one late at night, with many threads of work open at once across product, recruiting and writing. The agents run every day, and most of what they process is cached rereads of earlier context.

At that volume, keeping the agents working takes more of my attention than building them did, and each rule above is a line an IT department would write.

  • credentials
    Every machine counts as headless, and agents never ask for a machine password.
  • upgrades
    When the models behind the agents change, the work keeps going.
  • quality
    A helper agent gets work only while its quality holds.
  • budget
    Every agent has a budget that pauses it at the limit and a cap on paid searches per research run.
  • setup
    Agents may configure the new agents they set up.
  • help desk
    A question that needs me goes straight to the agent that owns that work.
The rules I set for my own agents, each one a line an IT department would write.

The department is already forming

Other companies are reaching the same conclusion in public.

Supabase is hiring a Head of AI Native Operations to stand up a new central team that builds the platform for agents and runs it in production. The post reads like an IT job description: least-privilege credentials, audit logs, every run traced with what it did, why and what it cost, and evals that say within hours of a new model's release whether to adopt it. The goal is 20 or more agents in production within six months, each with a named owner, so that other teams can ship agents of their own.

Moderna went further. In 2025 it merged its technology and HR departments under one chief people and digital technology officer, to plan the work across people and technology together. Jensen Huang described the same pairing at CES in January 2025, when he said the IT department of every company would become the HR department of AI agents.

The numbers say this will not stay small. Gartner expects the average global Fortune 500 company to go from fewer than 15 agents in use in 2025 to more than 150,000 by 2028, and only 13% of organizations think they have the right agent governance today (Gartner, Apr 2026). Its six steps for handling that sprawl include a central inventory of agents and managing each agent's identity, permissions and life cycle, which reads like an IT charter. IDC expects the agents organizations run to reach 2.5 billion by 2030, nearly 80 times the 2025 level (IDC, Aug 2026).

2025: fewerthan 15 agents2028: more than 150,000one square = 1,000 agents; the speck in the outlinedsquare is the 2025 count, drawn to scale13% think theyhave the rightagent governanceGartner, Apr 2026

The money is moving ahead of the plans. Companies planned to grow their AI spending 1.9 times this year and spent 4.7 times last year's level on average, because agents work on their own, run in parallel and keep going after people step away (Menlo Ventures, Nov 2026). Gartner expects the platforms that manage agents to capture over 60% of AI's compounded value by 2030 (Gartner, Mar 2026).

It happened once before

The historian Thomas Haigh traced how companies took in their first computers in the 1950s (The Chromium-Plated Tabulator). A new institution formed inside them: the data processing department, built out of the existing tabulating and systems-and-procedures teams. By 1958 thousands of computers were installed, and that department's shape had settled into one that changed slowly for the next twenty years.

Two things from that history carry over. Companies staffed the new department from the teams that already ran the work. IBM's Thomas Watson Jr. told punched card supervisors in 1954 that they had the best background for the electronic office. And in Haigh's reading, what marked the professional in that field was a manager's orientation more than technical excellence.

Agent management will go the same way. The people who know how a company's work gets done are the ones best placed to run the agents that do it.

IT never built its own computers

The IT department picked the vendors, wired the systems together and issued the machines, and each person made theirs their own. Good IT swaps the technology underneath while people barely notice, and the whole system gets better each time.

Agents need the same craft. Most engineers are composition engineers now. The job is picking the right models, vendors and open source tools for each task, stitching them into a workflow that runs in any host, and building the custom software and UI each job needs on top. Composing them well is closer to an art than a science. Everyone ends up with a setup shaped to them, held together by connective tissue everyone shares.

That gives agents two layers, the way devices always have. The company layer is shared: its memory and its connections to the tools it runs on. The person layer is how each person gets their work done, and each person shapes their own. The company issues the agent, and the person makes it theirs.

The host stays the person's choice. People move between Claude, ChatGPT and Grok, and an agent should show up inside whichever one they use with its playbook, memory, tools and app intact. When a new model is better, every agent built this way gets better with it, and the people using them shouldn't have to notice the swap.

The job is translating the work

The person in this role understands the pieces well enough to choose between them. Their main skill is translation. They take how a team works and turn it into playbooks, permissions and tools an agent can run, and they leave the team's work to the team and its agents.

That role matters more as the buying decision changes. Companies used to hire a person for a job. Software made them ask whether a tool could do it instead. Now the choice is a person, a tool or an agent, and an agent worth choosing covers the many things a job title does. Someone has to decide which jobs get one and keep each agent doing its job well.

Where this goes

Large companies will build this department themselves, the way Supabase is starting to. Most companies won't staff one yet, above all non-technical teams that just raised money or grew fast.

That is what Rails is for. Rails is a managed agents platform. Rails doesn't decide how work splits into agents. A company or a person does: one agent for a whole job, or one job split across several agents to lighten the mental load. Each agent has its own playbook, memory, tools and app, and Rails keeps it working as the models underneath change. Companies bring their own model keys and subscriptions. For teams without the department, we act as one, part time. We start by understanding how the team works, then set up its agents, including the custom software and UI each job needs, and stay on to run and improve them against the goals the team sets.

I already run my own agents this way. Every agent has a budget, granted tools, keys it never sees and a record of every call it makes, and everything that needs me comes to one place.