vaathi.dev

Agentic workflows (the loop)

Goal → plan → tools → observe → loop — vs a single chat reply.

Most chat use is single-shot: you ask, you get a reply, you’re done (or you ask again yourself). Agentic systems try to chase a goal by planning, calling tools, reading results, and looping until they finish — or until a limit stops them.

Single-shot first

DiagramSingle-shot chat — one ask, one reply (no tool loop)
Single-shot chat workflowThree boxes in a line: You write a prompt, the chat model answers once, you receive a reply. One arrow forward; no loop back.YouWrite a promptChat modelOne generationReplyDone (for now)

That’s enough for many beginner tasks. Your personal AI workflow sits on top: you still research, verify, and cite.

The agentic loop

DiagramAgentic workflow — plan, act with tools, observe, then loop or finish
Agentic workflow loopFlow starts at Goal, goes to Plan, then Act with tools, then Observe. From Observe a loop arrow returns to Plan if more work is needed. A separate path from Observe goes to Finish when the goal is met. Step numbers 1 through 5 label Goal, Plan, Act, Observe, and Finish.1Goalstart2Plan3ActTools / actions4ObserveRead resultsloop if not done5Finishgoal met →

Solid arrows = forward steps (numbered). Dashed teal arrow = “loop if not done,” also labeled in the diagram.

In plain words:

  1. Goal — What should be true when we’re done?
  2. Plan — Break the goal into steps (the system may rewrite the plan as it goes).
  3. Act — Call tools: search, code, calendar, files, APIs.
  4. Observe — Read what the tool returned; notice errors.
  5. Loop or finish — Update the plan and act again, or stop when the goal looks met.

See also: Glossary — Agent, Agentic workflow.

Why beginners go slow here

  • More steps means more ways to spend money, leak data, or take a wrong action.
  • A fluent loop is not the same as a correct loop — you still verify outcomes.
  • Prefer supervised / sandboxed agents before anything that can email, buy, or delete.

From the tool landscape: start with chat; treat agents as “powerful; start supervised.” For how apps reach cloud-hosted models in that loop, see how clouds serve frontier models.

Takeaway

Single-shot = one generation. Agentic = goal-driven loop with tools. Same models often power both — the workflow shape is what changes.

Try this

Pick a task (e.g. “find three reputable sources on X and draft a paragraph”). Sketch it once as single-shot chat, once as the five-step loop. Which steps would you insist on doing yourself?