Cron Jobs vs. Agentic AI

Here's the full picture:

Agents are just configuration files. Most commonly YAML, JSON, or Python code. They contain four things:

• Goal— Natural language description of what you want ("backup database daily")
• Tools — List of available functions (shell commands, HTTP calls, file I/O, etc.)
• LLM config — Where and how to reach the LLM
  • api_key: Your authentication token (e.g., sk-ant-...)
  • endpoint: The URL (e.g., https://api.anthropic.com/v1/messages)
  • model: Which model to use (e.g., claude-sonnet-4-6)
  • Other params like temperature, max_tokens
• Loop logic — The reasoning → action → observation → feedback cycle

How the agent knows where the LLM is

The api_key and endpoint URL are hardcoded in the config file (or loaded from environment variables for security). At runtime, when the agent needs to reason, it:

1. Prepares a request with the current context
2. Adds an Authorization: Bearer sk-ant-... header
3. Makes an HTTPS POST to the endpoint
4. The LLM server verifies the key, processes it, returns a response
5. Agent parses the response and continues the loop

That's it. The agent doesn't "learn"—it's stateless. Every time it runs (or starts a new session), it loads the config file fresh and starts over. If you want true learning, you'd need to:

• Save successful agent trajectories to a dataset
• Fine-tune the LLM offline on those trajectories
• Deploy the updated LLM to the agent config

The agent itself is remarkably simple: glorified orchestration layer is fair. But that orchestration—deciding what to do based on observations and reasoning—is exactly what cron jobs can't do.