We shipped an AI agent and it broke in ways nobody warned us about

A few weeks ago we pushed our first "real" AI agent into production — nothing fancy, just something that was supposed to triage support tickets and pull context from three internal systems. It worked great in the demo. Then it hit real traffic and started doing weird things: retrying the same failed API call in a loop, silently dropping context between steps, occasionally hallucinating a customer's account status because an upstream call had timed out.

We assumed we'd built something broken. Turns out we're far from alone — G2's State of AI Agent Builders 2026 report surveyed 7 vendors and analyzed 770 verified reviews, and the pattern it found matches almost exactly what we hit.

The single most interesting stat: 6 out of 7 vendors named API and system integration failures as the most common cause of agent workflow failures — not model quality, not prompting, not reasoning. Integration. The plumbing.

Orchestration came up just as much. Vendors said their orchestration layer is doing an average of 3.4 out of 5 possible jobs — task routing, retries, timeout handling, escalation logic, governance — all bolted onto what looks like "just a loop that calls an LLM" from the outside. One quote from a SnapLogic engineer stuck with me:

"Individual agents are relatively straightforward to build. The hard engineering is in the orchestrator: routing decisions, conditional branching, parallelization, retry logic, timeout handling, and deciding when to escalate to a human vs. retry. That logic is bespoke, it's brittle, and it's almost never documented properly."

That's exactly the code nobody wants to own six months later.

The report also pushes back hard on the "plug and play" pitch. Models drift, edge cases pile up, and if you don't have monitoring in place, quality degrades quietly until someone notices in a support escalation. One vendor put it bluntly: teams don't invest in the eval infrastructure and continuous monitoring needed to catch regressions before customers do.

None of this means agents aren't worth building — the same report found 43.5% of buyers write meaningfully more about what they like than what they dislike, and integration capability itself ranks in their top 5 most-valued features. The takeaway isn't "don't do this," it's "the orchestrator and the integration layer are the actual product, budget your time accordingly."

If you're building or evaluating agent tooling right now, it's worth a read: https://learn.g2.com/state-of-ai-agent-builders-2026