Orqisai
Turn any API into an AI agent. Safely.
Start new thread
safa

What’s the biggest challenge you’ve faced using AI agents in production?

by•

I’ve been experimenting a lot with AI agents connected to APIs recently.

They’re powerful, but I keep running into issues like:

  • lack of control over what gets executed

  • difficulty handling auth safely

  • limited visibility into what actually happened

Curious — for those building with agents, what’s been the hardest part in real-world use?

Would love to hear your experience.

11 views

Add a comment

Replies

Best
Alex Rotar
Visibility is absolutely the biggest bottleneck right now. When agents start chaining actions, debugging a hallucinated API call or an auth loop feels like finding a needle in a haystack. Building strict 'human-in-the-loop' checkpoints for anything executing a state change has been the only way we've managed to keep things stable. Curious what stack you’re using for the agent orchestration?
safa

@arotar Totally agree, visibility becomes the real problem once agents go beyond single calls.

We ran into the same issue: once actions are chained, a single bad decision can cascade and it’s almost impossible to trace back what actually happened.

The “human-in-the-loop” approach makes a lot of sense, especially for state-changing operations. We’ve been thinking along similar lines, but also trying to make the execution itself more observable by default, so every step is traceable without having to reconstruct it after the fact.

For orchestration, we started experimenting with LangChain early on, but a lot of the work ended up being around adding control and safety layers on top.

Curious, are you handling tracing/logging yourself, or relying on something like LangSmith?