Beginner-Friendly Topic
Take your time - it's perfectly normal to re-read this topic 2-3 times. Try the interactive code editor below to run code yourself. Use the Q&A section to check your understanding before moving on.You've got this!
Limitations of Agentic Systems
Agentic AI systems are powerful but they have real limitations. Understanding these limitations helps you design safer, more reliable agent systems.
Key Limitations
- Reliability: agents can make incorrect decisions, especially when the goal is ambiguous or the data is incomplete
- Cost: each step in an agent loop typically incurs API costs. Long or inefficient loops can be expensive
- Latency: a multi-step agent workflow takes longer than a single model call. This can be too slow for real-time applications
- Error propagation: an incorrect decision in step 3 can corrupt all subsequent steps
- Tool dependency: if a tool or API is unavailable, the agent may be unable to continue
- Safety risks: agents with access to write actions can cause real harm if they act incorrectly
- Unpredictability: agents using language models can behave differently on the same input at different times
- Auditability: it can be difficult to understand exactly why an agent made a particular decision
Working Within Limitations
The limitations of agentic systems are not reasons to avoid them. They are design constraints to work within.
For reliability: add validation at every step. Test the agent thoroughly on diverse inputs.
For cost: set step limits. Monitor and log API calls. Optimise the agent loop.
For safety: restrict write-action permissions. Require human approval for high-risk actions.
For unpredictability: use lower temperature settings. Add output validation. Test with many examples.
This course covers all of these engineering practices in later modules.
Key Takeaways
- Agentic AI systems are powerful but they have real limitations.
- Reliability: agents can make incorrect decisions, especially when the goal is ambiguous or the data is incomplete
- Cost: each step in an agent loop typically incurs API costs. Long or inefficient loops can be expensive
- Latency: a multi-step agent workflow takes longer than a single model call. This can be too slow for real-time applications