What is PRIMER?

PRIMER is a framework for building AI agents that follow a fixed loop: the user asks something, the AI reasons and plans, invokes a function, the app monitors the result, and the AI explains the answer (or the app recovers from an error). It was designed to work with the OpenAI API (e.g. GPT-4) and is extensible—you add new functions as needed.

When you don’t have a big MCP setup

You can use PRIMER when you don’t have (or need) a full MCP setup—no MCP servers, no MCP client, no protocol. Same loop: your app holds the prompt, calls your own functions (e.g. in a Python dict), and feeds results back to the model. MCP is for when you want that same loop standardized across many clients and servers; PRIMER is the lightweight version inside a single app.

Watch the PRIMER walkthrough:

The six steps

Letter	Step name	What happens
P	prompt_input	User asks something.
R	reason_plan	AI reasons and decides what to do (e.g. which function to call).
I	invoke_action	The app calls the chosen function with the given input.
M	monitor_result	The app captures the function’s output and feeds it back.
E	explain_output	AI formats and returns the final answer to the user.
R	recover_error	If something fails, the app reports the error and the AI (or app) handles it.

The loop runs until the AI returns an explain_output (success) or the flow stops after recover_error.

The loop

1. User sends a request (prompt_input).
2. Model responds with a step (e.g. reason_plan or invoke_action).
3. If the step is invoke_action, the app runs the function and appends a monitor_result (or recover_error) to the conversation.
4. Model sees the result and either calls another tool, or returns explain_output and the loop ends.

Your app owns both the loop and the set of functions the model can call (e.g. get train status, get PNR status, get train number from name).

The prompt

The model is given a system prompt that spells out the six steps and the available functions. It is asked to respond with JSON so the app can parse each step (e.g. invoke_action with a function name and input). Here is the prompt we used:

Follow the PRIMER steps to handle user requests using the available functions.

Steps:
- prompt_input: Wait for user input.
- reason_plan: Decide which function to call.
- invoke_action: Call the chosen function with input.
- monitor_result: Capture and observe the output.
- explain_output: Format and return the final answer.
- recover_error: Handle any issues and respond clearly.

Functions:
1. getTrainStatus(train_no: str) - Live train running status.
2. getPNRStatus(pnr: str) - PNR ticket & seat status.
3. getTrainNumberMapping() - Get train number from name.

Example:
START
{ "type": "prompt_input", "prompt": "get train running status for train 14682" }
{ "type": "reason_plan", "plan": "Call getTrainStatus for 14682" }
{ "type": "invoke_action", "function": "getTrainStatus", "input": "14682" }
{ "type": "monitor_result", "result": "Train is at NEW DELHI, delayed by 10 mins." }
{ "type": "explain_output", "response": "Train 14682 is at NEW DELHI, delayed by 10 minutes." }

Example

For a query like “get train running status for Karnataka Express”:

• reason_plan: get the train number first, then get status.
• invoke_action: call something like getTrainNumberMapping, then getTrainStatus with that number.
• monitor_result: each function’s return value is fed back.
• explain_output: the model summarizes the answer for the user.

So PRIMER is a structured agent loop inside a single app: same steps every time, with your own functions and error handling.

The original write-up and code: AiAgent – PRIMER with ChatGPT. For MCP concepts, see MCP Core Concept.

Next: How MCP is masked PRIMER — prompt parts mapped to MCP, diagram, and implementation sketch.