Agent-001: The Echo Loop

🇬🇧 English

🇫🇷 Français

Every agent, no matter how capable, is built around the same shape: read an observation, decide on an action, repeat. Before adding an LLM, tools, or evals, it’s worth building that loop on its own so later stages have something concrete to extend.

The code

agents/agent-001-echo-loop/agent.py implements the loop with no dependencies:

def run(input_stream, output_stream):
    while True:
        output_stream.write("User> ")
        output_stream.flush()
        line = input_stream.readline()
        if not line:
            break
        line = line.rstrip("\n")
        if not line:
            break
        output_stream.write(f"Agent> {line}\n")

• Observation: a line read from input_stream.
• Action: writing that line back to output_stream.
• Loop: the while loop, which continues until EOF or an empty line.

Passing in input_stream and output_stream (rather than hardcoding sys.stdin/sys.stdout) is what makes this testable without spawning a process – the tests in test_agent.py just pass io.StringIO objects.

Running it

python agents/agent-001-echo-loop/agent.py

Type a line, press enter, see it echoed. Press enter on an empty line to stop.

What’s next

Agent-002 replaces the echo with a real LLM call, keeping the same read-act-repeat shape.