world-agent
Train and deploy generative UI agents using behavioral programming constraints, A2A protocol for agent interoperability, and pattern registry for composition. Use when working with agent training, trajectory generation, reward computation, A2A integration, or deploying models.
npx skills add majiayu000/claude-skill-registry --skill world-agent --agent claude-code
Same command for any agent — swap --agent for codex, cursor, copilot.
Weekly change comes from our own snapshots, not the repository page — it measures attention, not adoption.
What it does
Guides training and deploying generative UI agents that coordinate using behavioral programming constraints, supports inter-agent communication via an A2A protocol, and leverages a pattern registry for composition. It covers when to use for training agents, generating trajectories, computing rewards, integrating with other agents, indexing patterns, and deploying models.
How it works
Describes a world agent architecture where a neural policy (FunctionGemma + LoRA) proposes tool calls, while symbolic constraints (bThreads) block invalid actions before execution. Training uses stories as the world model and a pattern registry for reuse; rules enforce runtime safety and guide tool execution. The agent runs in a sandboxed environment for executing generated code, and employs a tool discovery and A2A protocol integration workflow. Code examples show how to instantiate the agent with tools and model, trigger actions, and use a pattern registry, tool discovery, and sandbox components. It includes references to workflows and scripts for training, trajectory generation, reward computation, evaluation, and report generation, and demonstrates integration with HuggingFace Inference Endpoints and sandbox/runtime components.
Key components referenced:
- World model: Stories, story tests, Pattern Registry
- Policy: FunctionGemma (270M) + LoRA
- Symbolic constraints: bThreads
- Tool APIs, pattern registry usage, tool discovery, code sandbox, A2A protocol components
- Code snippets showing how to initialize and trigger world agent usage
When to use it
Use when training agents with GRPO on Google Colab, generating training trajectories from stories, computing rewards from story execution, integrating with other agents via A2A protocol, indexing validated stories as reusable patterns, deploying models to HuggingFace Inference Endpoints, executing generated code in a sandbox, and progressively filtering tools by user intent.
What it can touch
The skill references tools and modules from plaited/agent such as useWorldAgent, createCoreTools, createToolRegistry, and a variety of training, sandbox, A2A, and pattern registry utilities. It also mentions integration with @huggingface/inference and @anthropic-ai/sandbox-runtime, Bun, and Colab for training. Specific code imports shown include:
- import { useWorldAgent, createCoreTools } from 'plaited/agent'
- import { createToolRegistry, createCoreTools } from 'plaited/agent'
- imports for constraints, training utilities, A2A protocol, pattern registry, tool discovery, code sandbox, and skill scripts from 'plaited/agent'
Caveats
License is Apache-2.0. Requires @huggingface/inference, @anthropic-ai/sandbox-runtime, Bun >= 1.2.9, and Google Colab for training. The skill notes this environment and tooling requirements, without asserting guaranteed outcomes.
# World Agent ## Purpose This skill provides guidance for training and deploying generative UI agents that use behavioral programming for coordination. The world agent generates templates by composing validated patterns, learning from story execution feedback. **Use this when:** - Training agents with GRPO on Google Colab - Generating training trajectories from stories - Computing rewards from story execution results - Integrating with other agents via A2A protocol - Indexing validated stories as reusable patterns - Deploying models to HuggingFace Inference Endpoints - Executing generated code in a sandboxed environment - Filtering tools progressively based on user intent ## World Model Architecture The name "world agent" reflects its alignment with **world model** concepts from reinforcement learning and cognitive science. The agent learns an internal representation of how UI generation works—not by memorizing templates, but by understanding the *dynamics* of template composition. ### Why "World Agent"? | Concept | Implementation | Purpose | |---------|----------------|---------| | **World Model** | Stories + story tests | Defines valid UI states and transitions | | **Policy** | F
- Purpose
- World Model Architecture
- Why "World Agent"?
- Neuro-Symbolic Integration
- Key Architectural Concept
- Quick Reference
- Package Exports
- A2A Protocol Integration
- Present as A2A Agent
- Consume A2A Agents
- Pattern Registry
- Tool Discovery
- Basic Usage
- With MCP/A2A Integration
Scaffold a story file with multiple intents bun scripts/scaffold-training-story.ts button --category Button \ bun scripts/generate-trajectories.ts src/templates --output trajectories.jsonl Scaffold evaluation assets in your project claude /create-world-agent-eval .claude/eval Full evaluation (baseline + agent) bun scripts/run-eval-suite.ts .claude/eval Compare results bun scripts/compare-baseline.ts .claude/eval --format markdown Generate report
What does the world-agent skill do?
Train and deploy generative UI agents using behavioral programming constraints, A2A protocol for agent interoperability, and pattern registry for composition. Use when working with agent training, trajectory generation, reward computation, A2A integration, or deploying models.
How do I install it?
Run `npx skills add majiayu000/claude-skill-registry --skill world-agent --agent claude-code` — it drops the skill into your project so the agent can pick it up. Swap the --agent value for codex, cursor or copilot if you use one of those.
Where does this skill come from?
From majiayu000/claude-skill-registry, a repository with 534 stars. We read it straight from the repository tree rather than a submitted listing, so what you see here is what is actually published.
Is a popular skill a good skill?
Not necessarily. Stars measure attention, not adoption — a repository can trend for a week and be abandoned. That is why we show the weekly change from our own snapshots next to the total, instead of a single flattering number.