Agent Builder
Watch an AI agent stop being a magic box. This playground runs a scripted planner-executor loop and lights it up node by node: the planner decomposes the task and picks an action, the executor calls a tool, the loop observes the result, and a decide step checks the stop condition and the step budget before looping again or finishing. What most explanations hand-wave as "the agent thinks" is really this small control-flow graph — and the step-budget ticker is the hard ceiling that catches a runaway loop before it burns real spend. Runs entirely in your browser (0 uploads, works offline).
How to use it
- Press Advance one step to move the cursor one node at a time, or Run the loop to jump to the finish.
- Watch each node light up: the current node is highlighted, and everything behind it stays lit as the loop progresses.
- Read the step budget ticker — it shows how many of the loop's nodes have fired against the ceiling.
- Open the X-ray panel to see the phase the agent is in (planner, executor, observe, decide, stop) and why a weak stop condition is the classic way agents loop forever.
What this clears up (the fundamentals)
- An agent is a loop, not an oracle — plan → call a tool → observe → decide → repeat. Naming the nodes turns "the AI figures it out" into a control-flow graph you can reason about and debug.
- Planner and executor are different jobs — the planner decides what to do next and emits a step, not an answer; the executor turns that step into a real tool call. Every pass through the executor spends budget.
- The stop condition is the whole game — the decide node is where the loop checks whether it is done. A weak stop condition is exactly how an agent spins forever and runs up a bill instead of returning.
- The step budget is a safety ceiling — a hard cap on iterations means a runaway loop fails visibly and cheaply here, instead of silently on a live invoice.
Where it's used
A top-of-funnel AI-engineering explainer — the fastest way to feel why "agentic" means a plan-act-observe loop with a stop condition, not a smarter prompt. It is a miatz build-lab concept playable: play the seeded loop here, then learn to build the state machine, the tool-call budget, and the stop-condition stress test yourself in the AIE-301 agent module.
FAQ
What is a planner-executor loop?
It is the core agent architecture: a planner decides the next action, an executor runs it (usually a tool call), the loop observes the result, and a decide step checks whether to stop or go again. Repeat until a stop condition or budget ends it.
What is a stop condition, and why does it matter?
A stop condition is the rule that halts the loop — goal reached, budget exhausted, or a step cap hit. It matters because it is the most common failure point: a weak one lets the agent loop forever, re-planning and re-calling tools until it burns its whole budget with nothing to show.
What is the step budget ticker?
A visible ceiling on how many loop iterations the agent may run. In real systems it is a token or step budget that forces a runaway loop to fail fast and cheap. Here it fills as each node fires so you can see the loop pressing against its limit.
Does this call a real AI model?
No. This is a seeded, scripted run — the nodes and their outputs are baked into the engine so the loop is deterministic and completely offline. It teaches the control flow, not a live inference.
Is anything uploaded?
No. The whole loop runs in your browser from a fixed script — nothing is transmitted, stored, or logged. Turn off your Wi-Fi and it still works.
Limits
This is a seeded demo: a fixed, scripted planner-executor loop with hand-authored nodes, built to teach the control flow — it is not running a real model, real tools, or a real budget. Live BYOK inference (bring your own key, real tool calls, a real cost meter, and the five adversarial loop-inducing tasks) is a separate future mode. What the demo shows about the shape — plan, act, observe, decide, stop — is exactly how real agent loops work.
Related
Part of the miatz playgrounds. Explore the rest from the playground catalog.
Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.