LEARN. PRACTICE. GROW.Concept and mechanism
A generative model produces content conditioned on the request and context. Tokens are processing units that do not necessarily correspond to whole words. Embeddings represent content numerically and support semantic-similarity comparisons; similarity does not establish truth. The context window limits material available in one call. Context engineering organizes instructions, retrieved information, history, and tool outputs for the task. Redundant text may increase cost and make relevant information harder to identify. Compare models using task quality, modalities, latency, cost, constraints, and availability. A fluent answer can contain invented facts; confident presentation is not calibrated confidence. Define when to clarify, seek evidence, or refer to a person.
Guided application
A support assistant may summarize tickets; an agent may also select tools, inspect state, and perform authorized steps. MCP helps connect tools and systems to agents but does not grant business authorization. Memory retains context across interactions according to configuration and policy; it must not mix different users’ information. Multiple agents can divide work while adding coordination, cost, and failure points. AgentCore offers execution, identity, tool, and observability components; the application still defines boundaries and oversight. In a fictional calculation, 100 requests each use 2,000 input and 500 output tokens. With invented rates of 1 and 4 units per million, cost is 0.20 + 0.20 = 0.40 units before other services and retries. Also track task completion and satisfaction: low cost without a useful outcome can still be waste.
A summary does not necessarily require an agent with permissions to change production.
Common pitfalls
Embeddings treated as factual proof; MCP treated as authorization; unbounded shared memory; cost counting only input tokens.
Related topics: AI and ML: problem, data, and metrics · Foundation models: RAG, prompts, and evaluation · Responsible AI and explainability
Choose capabilities proportionate to the task and measure quality, cost, and control.
Reference: AIF-C01 domain2: Fundamentals of GenAI · AIF-C01