A founder note on the AI build-it-yourself trap
A founder asked me last week if she should build her own agent orchestrator. I told her no. Three engineering quarters at $300,000 per US senior developer is most of a million dollars of payroll, and that's the cheap version.
I have this conversation every week: "AI agents are so capable now we could build our own agent orchestrator. How hard can it be?"
It's the classic founder trap. You lose focus. The team you needed shipping product starts shipping infrastructure. The "two-week sprint" turns into a year. Good founders know what to leave alone. Orchestration is one of those things.
Why orchestration is where founders get burned
Without orchestration, your AI spend ends up on the AWS bill instead of in the product. The companies trying to scale agentic coding without the orchestration layer are showing up in headlines.
Uber burned through its entire 2026 AI budget by April. Four months. Their CTO told The Information he's "back to the drawing board." Uber has a $3.4 billion R&D budget and a full finance team.
A few weeks later, ServiceNow hit the same wall. They're the second major public company to disclose they blew through a full-year AI budget in the first months of 2026.
Per KPMG's quarterly AI survey, U.S. organisations are projecting average AI spending of $207 million over the next 12 months. Almost double their projection from a year ago.
The cost curve is moving against you. Anthropic shifted its tokenizer this spring; bills climbed up to 27% with no visible price change because the same input now generates 35% more tokens. On May 14, Anthropic split Claude subscriptions into an interactive pool and a separate agentic pool. OpenAI announced enterprise incentives the same day. None of this is in your favor.
What real orchestration actually has to do
If you sat down today to build a serious agent orchestration system, you'd need most of this before it could run safely inside a growing company:
- Role-based access control by repo, by agent, by model.
- Plan and diff approval gates on every agent run.
- Token budgets per task, per role, per team, with forecasts before execution and hard caps before runaway loops.
- Logs of every plan, tool call, and approval in your own database, queryable from the BI tools you already run.
- Persistent memory so corrections stick.
- Retries, checkpoints, and fallbacks when agents fail.
- Dashboards that show what's running and what it's costing.
- Custom promptware tuned to your domain, your security policies, your review style.
Plus prompt-injection guardrails, sandboxed plans, audit trails, and EU data residency if you sell into regulated industries.
That's a multi-quarter project at best. The bigger cost is the product feature you didn't ship because you were busy building plumbing.
So as founder to founder
Ivy's agent orchestration runs on cloud servers and spends the credits you already have: Azure from Microsoft for Startups, Anthropic startup tokens, AWS Activate balance. Those credits are engineering capacity you've already paid for. They just need an orchestrator pointing them at work.
Our startup package is Ship like a 50-person team: open-source orchestration installed on your existing Azure/Anthropic/AWS credits, with Ivy engineers handling the implementation. The engagement scales with what your team needs.
If you've already started building, talk to us before you commit another quarter. renco@ivy.app.
