Everyone in AI research is now saying agents need systems-level thinking, not bigger models.

They're right. With one exception.

Coding agents work differently from every other kind of agent — and confusing the two is costing teams six months.

Why Coding Agents Are Genuinely Model-Centric

Software engineering has structural advantages that almost no other business domain shares.

Built-in correctness feedback. Tests pass or fail. There's binary, unambiguous signal about whether the agent succeeded — and that signal is cheap to compute. The agent doesn't need a human reviewer to tell it whether the code works. The test suite does.

Self-documenting artifacts. Code is readable. An agent can review its own output — look at what it wrote, run the tests, see what failed, and iterate. This is rare. Most agent outputs (documents, decisions, data extractions) require external validation.

Minimal tool surface. Read a file. Write a file. Run a command. That covers the vast majority of software engineering tasks. Three tools. Compared to business process automation, which might need API calls to ten different systems, the tool surface is tiny — and each tool adds failure modes.

Natural iteration loops. Write → test → fix → test is a closed loop. The feedback arrives fast, it's interpretable, and the fix is local. Agents can run this loop hundreds of times in minutes.

These aren't advantages you engineer into a coding agent. They're inherent properties of software engineering as a domain.

Why Every Other Domain Is Different

Document processing doesn't have binary correctness. "Is this summary accurate?" requires human judgment.

Customer operations doesn't either. "Did the agent handle this conversation well?" involves context, tone, relationship history — none of which can be evaluated with a test suite.

Knowledge work, business process automation, data analysis, compliance workflows — most of what enterprises actually want to automate doesn't have the built-in correctness properties that make coding agents reliable.

For these domains, the lever isn't model quality. It's system design.

Better tool contracts — precise, minimal, well-documented interfaces — matter more than model capability. Tighter environment constraints — controlled inputs, structured outputs — matter more. Clear error budgets and human-in-the-loop gates matter more than whether you're on Claude 4.6 or 4.7.

What This Means for Your Strategy

Eighty percent solve rates on SWE-bench aren't evidence that AI is getting smarter at tasks broadly. They're evidence that software engineering is unusually amenable to AI automation.

The lesson from coding agents isn't "upgrade your model." It's "find the domain equivalent of a test suite."

For most domains, that means redesigning the task itself — building in correctness signals, constraining the tool surface, creating fast feedback loops — before you select a model.

The systems camp is right. Coding agents are the exception that proves the rule. And treating them as a template for everything else is the most expensive mistake I'm watching teams make right now.

Before you deploy an agent, score it across five dimensions: trigger clarity, output verifiability, error cost, tool surface, and iteration speed. Get the one-pager →