AI Agent Lab
A hundred and fourteen pages on AI agents, and every one of them runs. A real small language model works on your own GPU, real tools really execute, and every thought, tool call and mistake is printed on screen. Seventeen pages teach the mechanics, 73 walk through specific jobs agents do at work, and 23 cover the patterns behind them. No API key, no signup, nothing uploaded, and every page still works if you never load a model at all.
A five word definition
An agent is a model, tools, and a loop. It gets a goal instead of a message, it can call real functions to find things out or get things done, and it keeps going until the goal is met or a limit stops it. Everything else in this track is detail on those three words.
See one run
Start with What is an agent and Tool calling. Ninety percent of the idea is in those two pages.
Make it coordinate
Then Memory, Planning, Orchestrator agents, and Routing. This is where systems get real.
Make it safe to ship
Testing, Guardrails, Observability, and Monitoring. Skip these and your agent has a demo, not a job.
Two catalogs
73 agent jobs
One page per job an agent actually does at work, from ticket triage to invoice matching to incident response. Each has a live demo, the decision rules written out, how it fails, and the autonomy level it should start at.
Browse the jobs → 🧩23 patterns
The techniques the jobs keep reaching for: reflection, voting, verification, supervisors, budgets, retries, caching, evaluation, rollout. Each with a demo and an honest account of what it costs.
Browse the patterns →Start here: the seventeen core pages
What is an agent?
The six parts of every agent, then the same question answered by a plain model and by an agent side by side.
Start here. 6 minutes. → 🔧Tool calling
Run the seven tools yourself, then watch a model choose one, pass arguments, and recover from a bad call.
The single most important mechanic. → 🧠Memory
Scratchpad, conversation, long term, retrieval. Toggle memory off mid-conversation and watch the agent lose you.
Explains most agent bugs. → 📋Planning
ReAct or plan first? Give it a goal, get a numbered task list, then watch each task run with real tools.
For goals too big for one step. → 👥Multiple agents
The five patterns worth knowing, plus a live Drafter to Critic to Editor pipeline on your own text.
When one agent is not enough. → 🚄Routing and handoffs
A support ticket router that classifies, hands off to a specialist, and escalates to a human below a threshold you set.
The shape most working systems take. → 🎼Orchestrator agents
Give it a goal with three parts. It splits the goal, hands each task to the worker whose tools fit, runs them together, and merges the answers.
The shape of every agent team. → 🧪Testing agents
Eight real cases with assertions on the trace. Run them twice with a model loaded and watch the flake appear.
The part everyone skips. → 🧰Testing frameworks
Build a test by ticking boxes, run it against a live agent, then read the same test as pytest, promptfoo, DeepEval and vitest.
Choose a framework in ten minutes. → 🔭Observability
Traces, logs and metrics. Run one agent and watch it emit all three, then untick a signal and see which questions you can no longer answer.
The three signals, in one run. → 🌊Tracing
Read three recorded span waterfalls, click any span for its attributes, record a real one, and watch flat-rate sampling drop the runs you needed.
Find the widest bar. → 📜Logging
Build the log event field by field, run a real redaction pass on your own text, and price what full payload logging would cost at your traffic.
The evidence, and the bill. → 📏Metrics
Build a metric set and score it for coverage, watch the average hide the tail, turn a metric into an error budget, and see one label explode your cardinality.
Six numbers, chosen on purpose. → 📈Monitoring agents
A live production dashboard you can break on purpose. Cause a tool outage or a prompt regression and watch which alert fires first.
What happens after it ships. → 🛡Guardrails
A poisoned document tries to hijack the agent. Run it with the guardrails off, then on, and see what actually stops it.
Read before anything goes live. → 💼What agents can automate
The quick view: 57 automations in one filterable table with tools, autonomy and risk. Each has a full page in the jobs catalog.
Find your first use case. → 🛠Build your own
Name it, instruct it, choose its tools, set a step budget, run it, and export the configuration as JSON.
Put it all together. →The six things an agent can actually do
Find out
Search, retrieve, read a document, query a database. Most useful agents are mainly this.
Compute
Arithmetic, dates, conversions, aggregation. Never leave these to the model itself.
Decide
Classify, route, prioritize, choose the next step. Cheap, and it is where agents pay off first.
Create
Draft, rewrite, summarize, translate, restructure. The part everyone thinks of, and the least dangerous.
Act
Write to a system: file a ticket, update a record, send a message. Everything risky lives here.
Check
Verify a claim against a source, run assertions, flag what does not hold. The step that makes the rest trustworthy.
What you need
| To read and use every page | Any browser. Nothing to install, nothing to sign up for. |
|---|---|
| To see real model output | Chrome or Edge 113+ on desktop or Android, or Safari 18+, so WebGPU is available. The model downloads once (0.4 GB to 1 GB) and is then cached. |
| To keep your data private | Nothing you type on these pages is sent anywhere. The model runs on your machine, and the tools are JavaScript running in your tab. |
| Cost | Zero. No API key exists on this site, which is also why the model is small and sometimes wrong. That turns out to be useful: you can watch it fail and see why the checks matter. |
Free from AI School - no signup, everything runs in your browser.