Playground
The simplest, fastest way to build AI agents.
Every common agent pattern, one page each. Watch it run step by step, see the CriogenAI code that builds it, and change the task. No sign-up and no key needed. Add your own key if you want a real model to make the decisions; it stays in your browser.
“Fastest” is measured: CriogenAI against LangGraph and CrewAI on the same work, in the benchmark lab.
Reasoning
One agent, different ways to think.
- ReActThink, act with a tool, observe, repeat.ready-made in CriogenAI
- Chain of ThoughtReason step by step in one call, then answer.ready-made in CriogenAI
- Tree of ThoughtsBranch into several ideas, score them, grow the best.built from CriogenAI primitives
- Plan-and-SolveWrite a plan first, then carry out each step.built from CriogenAI primitives
- ReWOOPlan every tool call up front; run them; solve once.built from CriogenAI primitives
- ReflexionTry, test, write a lesson, try again with it.ready-made in CriogenAI
Multi-agent
Several agents, different ways to share the work.
- RouterOne call picks the specialist; only that one runs.built from CriogenAI primitives
- SupervisorA coordinator calls workers in turn until done.ready-made in CriogenAI
- Hierarchical teamsA manager splits work across team leads, who delegate.ready-made in CriogenAI
- Multi-agent collaborationSpecialists pass the work along a pipeline.ready-made in CriogenAI
- DebateTwo agents argue opposite sides; a judge decides.built from CriogenAI primitives
- Parallel fan-outSeveral agents work at once; one combines.ready-made in CriogenAI
- SwarmAgents hand off to each other; no one is in charge.ready-made in CriogenAI
- Human in the loopThe agent drafts; a person approves before anything happens.ready-made in CriogenAI
IBM agent types
The five classic types, from IBM's guide to AI agents. See IBM: types of AI agents.
- Simple reflex agentIf this, then that: rules on the current reading only.no model needed
- Model-based reflex agentRules plus an internal picture of what it cannot see.no model needed
- Goal-based agentSearches for actions that reach a goal, and replans.no model needed
- Utility-based agentScores every option by how good it is, picks the best.built from CriogenAI primitives
- Learning agentA critic scores its actions; it gets better with experience.no model needed
About your key
- It is optional. Without one, every pattern runs with scripted model replies.
- It is kept in this tab's memory only. It is never saved and never sent to this site; reload the page and it is gone. “Forget key” drops it at once.
- Your browser sends requests straight to the provider you chose (OpenAI, Anthropic, Google or OpenRouter). Each run makes at most 16 short calls.
- Tools are always simulated: nothing is booked, sent, refunded or deleted.