Prompt engineering or agent engineering: where is the problem?
Improving an instruction and designing a system that acts are related jobs. Identify the layer where a failure actually occurs.
A prompt can request a document summary. An agent-based system can find the document, check its version, summarize it and prepare an update elsewhere. The second case adds actions and state. Improving the opening sentence does not, by itself, address all of those responsibilities.
The architectural distinction described by Anthropic separates predefined workflows from systems in which a model dynamically selects its next steps. The article is a conceptual reference from 2024, not a current product catalog. Here it helps identify the structure a task requires.
What an instruction can clarify
Prompt engineering organizes the objective, context, constraints and response format. If a summary must preserve deadlines and owners, that requirement should be explicit. Examples can demonstrate the intended level of detail. Evaluation compares the answer with its source material and the stated criteria.
When the supplied material is sufficient and the output requires no action in another system, a well-structured call may meet the need. Adding memory, tools and continuous execution unnecessarily introduces more components to test and maintain. Complexity should solve an identified problem rather than signal that the application is sophisticated.
What changes when the system can act
Actions require tool definitions, permissions, error handling and stopping conditions. The system must distinguish a failed lookup from a successful lookup with no results. It also needs to avoid repeating an externally visible operation simply because a clear confirmation was not received.
Consider a hypothetical agent preparing an article update. Reading the text and proposing a revision is one task. Publication requires identifying the correct record, preserving existing fields and checking whether someone else edited the article after it was read. A prompt telling the agent to be careful does not implement that concurrency protection.
The tool interface is part of the solution
Anthropic’s article on tools for agents emphasizes clear descriptions and evaluation of tool use. The practical task is to make inputs and effects understandable so the model does not have to infer a hidden rule from an ambiguous function name.
An update function with several unexplained fields leaves uncertainty about what will be replaced. An interface identifying the record, expected revision and changed fields reduces that ambiguity. The server should validate its contract and return a recognizable error when the condition is not met. Written guidance complements that behavior; it cannot stand in for its implementation.
Investigate the layer that actually failed
If an answer omitted information present in its source, examine context and instructions. If a tool returned incomplete data, inspect retrieval and pagination. If an update overwrote another person’s work, inspect version checks and mutation behavior. These failures can appear in one conversation while requiring different repairs.
Record representative cases and repeat them after changing a prompt, tool or model. The agent governance article covers permitted effects. The method for using AI on web projects organizes verifiable deliveries. The productive question is not which discipline replaces the other, but which part of the system must change for the work to succeed and remain explainable afterward.
About the author
Tiago F SantiagoComments
No comments yet
Share a question or an experience related to the article.
Related articles
