What is an agent?#
In the context of large language models, an “agent” is an LLM-based system that can autonomously carry out tasks, make decisions, and interact with its environment. It combines the reasoning power of an LLM with the autonomy of a traditional software agent. A few defining traits:
- Autonomy — an agent can understand a task, plan, and act on it without constant human intervention.
- Goal-directed — it’s given a specific objective and works toward it.
- Environment interaction — it can perceive its surroundings and adjust its behaviour as conditions change.
- Reasoning and decision-making — it uses the underlying LLM’s capabilities to reason through non-trivial decisions.
- Tool use — many agents are built to call out to external tools and APIs: search engines, databases, or other software.
- Continuous learning — some more advanced agents can learn and improve from experience.
- Multimodality — some agents can process and generate more than just text, including images and audio.
Common agent frameworks#
A quick comparison of a few frameworks I looked at, roughly by strengths and how steep the learning curve is: LangChain, AutoGPT, Hugging Face Transformers, NVIDIA NeMo, and Microsoft’s DeepSpeed Chat — each has a different sweet spot depending on whether you’re optimising for flexibility, ease of setup, or production performance.
A worked example: telecom customer service#
As a concrete example, I sketched out a customer-service agent for a telecom scenario: it takes speech input, converts it to text, looks up account information (like balance) in a database, generates a response, and converts that back to speech. One deliberate design choice worth calling out: if the model’s answer doesn’t explicitly mention the account balance, the system appends it automatically — partly for compliance and consistency, and partly so the answer is useful even if the model’s own response wanders off-topic.
