Three loops in a trench coat: the real anatomy of an agent

In plain terms

An "AI agent," dismantled, is three nested loops: one that talks to the model, one that executes the tools it demands, and one that waits for a human to say yes. The first is trivial, the third is an engineering hell. And the job of a developer is shifting towards writing these loops rather than writing the prompts.

The loop, not the prompt

The post from bobbytables.io does what too few "agentic" contents do: it shows the code. The inference loop is a nearly stupid while loop - you call the completion API, you stack the result in the history, and you exit as soon as the model no longer requests a tool. The tools loop takes over when it requests one: you resolve the name, you call the function, you return the result correlated by its tool_call_id.

This is where the trap lies that the author rightly points out: a tool call is also inferred text. Therefore, the function name and its parameters can be hallucinated. Your tool dispatcher is not a router, it's an attack and validation surface.

The third - the human loop - is not technically a loop. It's a blocking call: someone, somewhere, decides to continue or not. It's also the only one that breaks your process, because a human can take three days to respond. Hence the use of durable execution engines like Temporal.

Under the hood

while continueInferenceLoop
 inferred = aiClient.completeChat(messages: chatMessages)
 chatMessages.push(inferred result)
 if inferred.toolCalls.length == 0
 continueInferenceLoop = false
end

Four lines. Everything else - the persistence of the history, the idempotence of the tools, the recovery after human approval - is classic distributed systems engineering. Not AI.

So what

The industrial counterpart of this anatomy has a name, and The Pragmatic Engineer just documented it: loop engineering. Addy Osmani's formula sums up the shift - "You design the system that does it instead": you no longer formulate the request, you design the system that formulates it in a loop. Boris Cherny (Anthropic) goes further: "My job is to write loops". The technique descends from Geoffrey Huntley's "Ralph loop", a simple Bash loop that recalls the agent until the goal is reached.

The uses that have shipped are prosaic: PR automatically opened on a new Sentry ticket, fixing flaky tests during nightly runs, triaging alerts, long migrations. Since May 2026, Codex, Claude Code, and Hermes expose a /goal command that wraps all this up.

If you're hiring, the skill to look for is no longer "knows how to prompt". It's: knows how to make a loop idempotent, observable, and interruptible.