LLM Agents Demystified (github.com)

LLM agents are not a magical new model class but a system pattern: an autoregressive LLM is wrapped in a control loop that interprets observations, emits structured “actions” (e.g., API calls, database queries, or natural-language tool invocations), receives tool outputs (observations), updates state/memory (often via a retriever in RAG setups), and repeats until a task-complete signal. Common designs separate planning (high-level decomposition), execution (calling tools), and memory/retrieval for grounding; patterns like ReAct combine reasoning traces with tool calls to produce more interpretable, stepwise behavior. Implementation details that matter: action schema/serialization in the prompt, how observations are concatenated into the model context, retrieval architecture (dense vs sparse), and when to re-invoke the model versus using cached state. For the AI/ML community this matters because agents extend otherwise static LLMs into interactive systems that can perform multi-step, grounded tasks and access up-to-date or private data. That power brings clear trade-offs: fewer hallucinations when tools ground answers but higher latency, API-call cost, and attack surface from tool misuse; more brittle behavior if action formats or termination criteria are poorly specified. Key research and engineering challenges are robust tool specification, memory and retrieval strategies for long-horizon tasks, efficient multi-call reasoning (cost/latency), verification/debugging of agent traces, and safety/alignment when agents acquire persistent capabilities.