Large language models (LLMs) are next token prediction machines. They are good at generating phrases or content (image, audio and video) based on provided inputs. At the early stages of LLMs, the key to effective #AI usage was to provide the best prompts for the model to guess satisfactory outputs: Prompt Engineering. Manual prompting couldn’t handle enterprise workloads at scale (like answering to each request to the the customers service), so Agentic AI emerged. An AI agent is an #LLM wrapped with a set of tools (programs to query data or trigger actions in internal or external systems), a memory to carry a conversation and an orchestrator that makes the connection between the components. For an agent to succeed in autonomy, it must be able to provide the right amount of context along with every request to the LLM. It usually consist of:
- A System Prompt that states the LLM persona. Eg: “You are a sales manager providing quotes upon customer request”
- Available knowledge. Eg: “There will be a sales campaign for TVs during the month of may”, “We might face PCs shortage by next quarter due to memory crisis”
- Available tools. Eg: “You can query the inventory database for available products using our internal MCP (model context protocol) server” Depending on the range of topics the agent is expected to handle, the context can overflow, leading to token inflation and deterioration of LLM performance: we call it Context Rot. To overcome it, Context Engineering entered into play. It is the discipline of providing the right amount of context with each LLM request. It leverages techniques like Retrieval Augmented Generation (RAG) to select only the subset of available context that is relevant for a user’s request. Some problems remain here:
- A static system prompt cannot effectively match all user queries: A request for a new TV can be best handled by a sales expert while a complain about a dysfunctional device should be handled by an after-sales expert. This also applies to available tools.
- Some requests can be effectively handled by small model while others might need more reasoning models. Proper model selection is essential for optimal costs. Multi-AI Agents systems are typically built as specialized sub-agents with dedicated system prompt, available context, tools and selected LLM model. On top of that, we add a routing agent to effectively dispatch requests. It comes with many advantages:
- Separation of duties: Each sub-agent is equipped with only the necessary tools and knowledge for its specific domain
- Cost effectiveness: Each sub-agent is configured with the optimal LLM for his assignments While improving effective prompt and tools usage, multi-AI agents systems come with their own limitations. As you add more specialities (sub-agents), your multi-AI agents system can become complex and lead to a maintenance nightmare. It requires engineering skills to implement, deploy and maintain each new agent. The system/process can be improved by introducing the capability for domain experts to add new specialities to existing agent while keeping its prompt concise and effective: #Skills.
Skills are a catalog of specialities that an agent can activate when necessary to fulfill a user’s request. Skills are based on progressive disclosure: The catalog (including descriptions and metadata) of available skills are added by the agent to the initial system prompt, but the detailed instructions about how to perform any special task is loaded only when requested by the model (skill activation). This way, new capabilities can be added to the catalog without changing the agent’s code. A skills can also be accompanied with additional tools implemented as scripts, or refer to additional context that the model can request to load. Skills are portable, meaning that your company can use the same catalog with different agents or LLMs (Copilot, Claude, Gemini…). They are also easy to implement and maintain by your business experts, as all they require is specific directory per skill that includes a SKILL.md file. You can leverage skills to turn your company expertise into structured knowledge that AI can use to effectively automate your business. Depending on your industry, you can also integrate publicly available skills to your catalog for your agent to implement standards and best practices.
If you want to know more about agent skills and how they can be applied to your specific use case or industry, let me know in the comments.
Leave a Reply