# What's the biggest thing that makes AI agents unreliable in real-world workflows? — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/55c652cf7888c228

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-10-23 20:07 ingested 2026-10-24 02:35

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1xdolej/whats_the_biggest_thing_that_makes_ai_agents/)

## retained text



```
I've been reading discussions about Agent Skills, and one thing that stood out to me is that people aren't necessarily looking for more skills. They're looking for agents that actually work reliably.

Some people mentioned context retrieval, others talked about testing before implementation, breaking tasks into smaller steps, and helping agents recover from mistakes.

It got me thinking about what the real bottleneck is when building agentic workflows.

For those who've built agents beyond simple demos, what's caused the most problems for you?

1. Poor context or memory

2. Incorrect tool calls

3. Long workflows breaking midway

4. Agents repeating mistakes

5. Token costs and context limits

6. Something else entirely

I'd be interested in hearing about problems you've actually encountered, not just theoretical limitations.
```



document fetched 2026-10-24 10:22 · canonical https://reddit.com/r/AI_Agents/comments/1xdolej/whats_the_biggest_thing_that_makes_ai_agents
