source: kdnuggets: 5 key concepts behind agentic ai every engineer must understand
level: technical
an llm alone only generates text. to check a database, call an api, or send an email, you need tool use. the model context protocol (mcp), introduced by anthropic in 2024 and now under the linux foundation, standardizes how models discover and call tools via json-rpc. it adds token overhead, so for high-throughput pipelines, direct calls may be better. mcp shines when you need oauth, multi-tenancy, or non-engineer tool integration.
agents need memory because llm calls are stateless. modern systems extract facts into a vector database and retrieve relevant slices before each response. tools like mem0, zep, and letta handle this, with zep using temporal knowledge graphs for time-based queries. context engineering has replaced prompt engineering, focusing on selecting and compressing information rather than just filling the context window.
planning loops let agents decide, act, observe, and repeat. the react pattern from 2022 interleaves reasoning and actions. now, frameworks add retries, self-correction, and task decomposition. multi-agent orchestration splits work across specialized agents, coordinated by an orchestrator. protocols like a2a let agents from different frameworks collaborate. evaluation and observability are critical: tracing shows what happened, and evaluation judges quality. most agents fail to reach production, but those that do see high roi.
why it matters: understanding these concepts helps engineers build agentic systems that actually work in production, avoiding common failures and achieving practical returns.
source: kdnuggets: 5 key concepts behind agentic ai every engineer must understand