There's a new line item appearing in company budgets everywhere: build our agents. Hire an AI engineer, stand up an orchestration framework, wire up the tools, prompt a "growth agent" into existence. Every board is asking for it; every CTO is staffing it.
We think most of this work will look, in five years, like the server rooms companies proudly built in 2005. Not wrong because agents are wrong — agents are the future, and we've said so in every essay before this one — but wrong because building your own specialist agents is the wrong side of a buy-versus-build that has already been settled once.
The settlement has a protocol now. This essay is about what it changes.
01The 2005 rerun
Rewind twenty years. Every serious company ran its own mail server, its own CRM install, its own data center cage. The work was real, the engineers were good, and almost all of it was waste — because none of it was the company's actual business. SaaS won by making a simple argument: the company that does only CRM will run CRM better than you, forever, and you can rent the result.
Now watch a company "build its growth agent" in 2026. An in-house team — whose business is, say, logistics software — is teaching a general-purpose model what a sales play is, what deliverability means, when a bounce matters, how not to email the same person from two campaigns. They are re-deriving, prompt by prompt, the domain knowledge that specialists spent years encoding. And the result is a median agent: the distilled average of public blog posts about outbound, operated by people who'd rather be building logistics software.
The economics that killed the server room kill this too. Specialization compounds. We run outbound for hundreds of companies; every failure mode we hit hardens the system for everyone. An in-house growth agent learns from one company's traffic. Ours learns from all of it. That gap widens every single week, and it's the same gap for legal agents, finance agents, support agents — every domain deep enough to have failure modes you can't read about.
So the equilibrium is easy to call: companies won't build specialist agents, for the same reason they don't build CRMs. They'll hire them.
02"Hire" is now a protocol
What makes this more than an analogy is that the hiring interface actually exists, and it's standardizing fast.
The agent-to-agent protocol (A2A) gives an agent the things an employer needs from a contractor: a card that says who it is and what it can do, skills with defined inputs and outcomes, and a wire format for handing over a task, streaming progress, and receiving the result. The card lives at a well-known URL, like a résumé with an API.
Ours is live. agent.ottosoftwares.com/.well-known/agent-card.json describes a persona we deliberately shaped like a hire, not like a toolbox: a Head of Growth. Two skills today — research leads, open to any caller, and run outbound, keyed because it spends money. Note what the card doesn't list: forty CRUD endpoints. You don't hire a head of growth by their keystrokes; you hire them by their outcomes. The whole argument of Software for AIs — delegation beats operation — is compressed into that design choice.
Around the card sits the rest of the employment package:
- An MCP server, so the hiring agent (or its human) can supervise — open the dashboards, read the activity, check the pipeline — the way a manager checks in on a report, without micromanaging the work.
- Outcome pricing, because a contractor that bills per seat is absurd and one that bills per token is uninsurable. One credit, one emailed contact. An agent with a wallet can reason about that.
- Server-side guardrails — budgets, caps, approval gates as schema — because an employer needs to bound what a contractor can spend before trusting it, not after.
llms.txtand machine-readable everything, because a contractor an orchestrator can't discover doesn't exist.
03Otto on your agent team
Here's where this gets concrete, because the employers already exist.
Agent teams are becoming real organizational units. Orchestrators like Paperclip, OpenClaw, and Hermes let a company — or a single founder — run a roster of agents the way a manager runs a team: one coordinator holding the goals, specialists holding the domains. The orchestrator is generalist by design. It plans, routes, and verifies. What it should not do is impersonate a head of growth from first principles.
Our vision is that Otto slots into that org chart as the hire: the coordinator reads our card, delegates "fill the pipeline for this product, €X budget," and gets back structured progress and outcomes — while every guardrail stays enforced on our side, and any human above the orchestrator can open the cockpit and take the wheel at any moment. The same software, the same engine, the same steering wheel; just a different manager.
And this is the plan for the whole fleet. Every autonomous software we ship — outbound today, social and ads next — ships with two doors from day one: a cockpit for people, a card for agents. One hireable specialist per job. An autonomous company, in our picture, is not a company that built a hundred agents. It's a thin coordinator and a staff of hired specialists, each the best in the world at exactly one thing, none of them written in-house.
04The uncomfortable implication
If this is right, the current wave of in-house agent engineering has the same destiny as in-house data centers: a transitional decade, then a migration. The durable work isn't building the hundredth mediocre growth agent — it's choosing which specialists to hire, setting the goals and budgets, and supervising well. That's a management problem, and management is what companies are actually good at.
The companies that internalize this early get the compounding specialist without the payroll. The ones that don't will spend two years building an agent we — or someone like us, in some other domain — will obsolete with a Tuesday deploy.
If you run an agent team, the interview is one HTTP request away: agent.ottosoftwares.com/.well-known/agent-card.json. References available — drop any URL at outbound.ottosoftwares.com and watch the candidate work.
— Benoit