Skip to main content
Ask Claude a question and you get one pass: read the request, write the answer, done. That single pass is enough for most conversation, and it fails completely for real tasks, because real tasks contain surprises the first plan cannot see. An agent is the same model taken off the single pass and put in a loop. It plans a step, acts with a tool, observes what actually came back, and revises the plan before going around again. It stops when the goal is met, not when the writing is finished. That loop is the whole trick behind every agent product you will ever meet, and once you can see it turning, agents stop being mysterious.

​
Watch a loop run

One real task: reconcile a month of expenses. Step through Plan → Act → Observe → Revise and watch progress move. Then switch to the fuzzy goal and see the same loop stall.
Two moments deserve a look. First, the dip. Early in the sharp run, progress goes backward when the agent finds charges with no receipts. Finding a problem looks like regression and is actually the loop working. A one-pass draft would have simply never met those five charges. Second, the stall. In the fuzzy scenario the loop runs cleanly and goes nowhere, because “clean up the expenses” has no finish line. The loop did not fail. The goal never existed.

​
The four beats

  • Plan is cheap and disposable. Good agents plan one step deep, act, and let reality vote, rather than plotting ten steps into fog.
  • Act means a tool call: run code, search, open a file, hit a system. Acting is what separates an agent from a very confident narrator.
  • Observe is the beat that makes it a loop. The agent reads the actual result, including the failures. Skipped observation is how agents cheerfully build on a step that silently broke.
  • Revise closes the circle. The plan absorbs what observation found, and the next lap starts smarter than the last one.

​
Your three jobs

The loop runs itself. What it cannot supply is judgment about what is worth doing, and that stays with you:
1

Define done before it starts

“Done means a table finance can act on, one row per charge, mismatches flagged” gives the loop a wall to stop at. Without it, the agent either stops early or laps forever. Writing the finish line is the single highest-leverage sentence you contribute.
2

Set checkpoints proportional to stakes

For a throwaway analysis, let it run to the end. For anything that touches real systems or goes to real people, ask the agent to pause and show its plan, or its findings, before it commits. You are the human between the loop and consequences.
3

Inspect the lap log, not just the deliverable

The record of what the agent did, found, and changed course on is where trust is earned. A deliverable can look right by accident. A log that shows the five missing receipts being found, chased, and explained cannot.

​
Where you will meet the loop

You will rarely see the words plan, act, observe, revise on screen, but you will watch the loop turn in three places. Research runs it against sources: search, read, notice gaps, search again, until it can write with citations. Cowork runs it against a whole task while you are in a meeting: plan the steps, do them, check the result, file the output. Claude Code runs it against software: change, run, read the errors, change again. Same loop, different arenas. Which means the three jobs above are portable: define done, set checkpoints, read the log, everywhere an agent works for you.

Skills, MCP, connectors, and built-ins

The loop is the engine of agentic work. The vocabulary lesson covers the equipment it gets to act with.