Ordinary programs, repeating steps

Ordinary automation runs in loops.

Most software repeats: look at input, choose the next move, do it—then repeat or stop. Your program still runs tools and decides when to stop.

Programs may save results and enforce stopping and safety.

Play: sort one parcel

Sort one parcel twice

One parcel, same sticker and note every round. Route it twice—first with a rule card that reads the sticker only, then swap the chooser at Choose to read both signals.

Pick the simplest chooser

Clear, stable fields? Use rules. Conflicting notes or messy text? A simple agent can use a language model at Choose. Your program still checks every suggestion before it runs.

• Rules — clear fields, stable paths
• Hybrid — rules plus a language model where text is messy
• Language loop — open-ended work with many tools

Takeaway

Same loop. A language model often helps at Choose. Your program still owns Do—and saving, stopping, and safety.

Primary sources

• Russell & Norvig — Intelligent agents (AIMA Ch. 2)
• Yao et al. — ReAct: Synergizing Reasoning and Acting
• Anthropic — How tool use works
• Anthropic — Building effective agents

while not done:
  input = look(state)
  move = choose(input)
  if validate(move):
    state = run(move)