AI agents are easy to create with modern SDKs but present challenges like high costs, rate limiting, and security concerns.
A basic chatbot can be built from scratch using Node.js to read user input and make plain fetch calls to an LLM API like Mistral.
Enhance an agent's capabilities by defining a tool for shell access, sending its definition in the API payload, and executing commands locally.
Mitigate the risks of unrestricted shell access by running the agent's code execution within an isolated Docker sandbox environment.
An agent can be prompted to modify its own source code to add new tools, such as a memory function that saves information to a file.
Structure complex tasks by creating a main agent that delegates specific jobs, like coding, to specialized sub-agents via a tool call.
An AI agent's fundamental architecture consists of a model for reasoning and a harness, which is the environment and tools it uses for execution.
Because LLMs are stateless, every API call must resend the entire conversation history, causing input token costs to grow cumulatively.
A live demonstration shows how adding large tool outputs like file lists to the conversation history makes subsequent simple requests extremely expensive.
A real-world agent that processes emails and calendar events shows how a single user prompt can trigger multiple tool calls, resulting in over 100,000 input tokens.
Remember that LLMs are stateless, tool execution is local, and an agent is fundamentally a model combined with a harness of tools and an environment.
Jobs that call for the skills explored in this talk.