The hidden ops cost of custom-built agent scaffolding

Every team that builds its own agent scaffolding has the same experience: the first version takes two engineers three weeks and feels like a reasonable investment. Eighteen months later, two engineers spend 30% of their time maintaining it, and no one is sure they can safely upgrade it without breaking something in production.

The initial cost is visible. The ongoing cost is invisible until it isn't.

What "custom-built scaffolding" actually costs

The build cost

Call it 2 engineers × 3 weeks = ~6 engineer-weeks to build a reasonably solid agent scaffold. That includes state management, retry logic, tool registration, basic auth, and logging. Not fancy — just solid enough to ship.

At $12K/week fully-loaded cost per senior engineer, that's $72K of engineering labor. This is the number teams usually see and accept because it's amortized over many agents.

The maintenance cost (hidden)

Here's where the math breaks. Custom scaffolding doesn't depreciate — it accumulates debt:

• LLM provider API changes. When OpenAI changes function calling syntax, or Anthropic changes their tool use format, someone has to update the scaffolding. That's not the product roadmap. That's unplanned infrastructure work.
• Security patches. Dependencies in your scaffolding have CVEs. Someone has to track them and patch them. At a minimum, 2–4 patches per year, each requiring test and deploy cycles.
• New agent requirements. The second agent you build will have requirements the scaffolding wasn't designed for. You extend it. The third agent extends it further. Over time, you have a framework — one with no documentation, no test coverage, and no maintainer who fully understands it.
• Oncall incidents. Custom scaffolding is the most common root cause in agent production incidents. State persistence bug, retry storm, auth token expiry — these are scaffolding problems, not product problems. But they wake your engineers up at 2am.

Realistic ongoing cost: 0.5 engineer per quarter, full-time equivalent, on scaffolding maintenance. At $12K/week, that's $24K/quarter or $96K/year — more than the original build cost, every year, forever.

The opportunity cost (most hidden)

The costliest line item doesn't appear on any invoice. When your best engineers are maintaining scaffolding, they're not building the product features that differentiate you.

Scaffolding maintenance is often handed to whoever is available, which is usually the newest engineer on the team. That engineer spends their first months learning undocumented infrastructure instead of shipping product. Their onboarding time is longer. Their early contributions are slower.

Multiply that across every new engineer who joins, and the scaffolding has become a permanent drag on team velocity.

The calculation teams don't do

Most teams evaluate custom scaffolding by comparing build cost to the cost of a managed solution. That's the wrong comparison.

The right comparison is:

Build cost + Ongoing maintenance cost + Opportunity cost

versus

Managed solution cost + Configuration time

At the one-year mark, the custom scaffolding is almost always more expensive — even before accounting for opportunity cost. At the two-year mark, it's rarely even close.

When custom scaffolding is the right choice

There are legitimate reasons to build custom scaffolding:

• You have unique requirements no existing framework handles
• You have deep expertise in the problem space and a team committed to maintaining it long-term
• You're building a framework product and the scaffolding is part of your IP
• Regulatory constraints require full ownership and auditability of every layer

None of these are "we haven't looked at what's available" or "it'll be faster to just write it ourselves."

The total cost model

Before your next sprint planning where someone proposes building the agent scaffold internally, run this math:

• What will this cost to build?
• Who will maintain it when the builders are gone?
• What happens to the second and third agents that extend it?
• What incidents will it cause, and how many oncall hours will those cost?
• What product features won't get built because this got built instead?

The answers to those questions are the real cost of custom scaffolding. They're also why we built Agent Chassis — to make the standard frame a dependency you install, not infrastructure you own.