Emergent Capabilities

Agents can create new tools at runtime when no existing capability matches their need. Enable with emergent: true in your agent config.

Enable in agent.config.json

{
  "emergent": true,
  "emergentConfig": {
    "maxSessionTools": 10,
    "sandboxTimeoutMs": 5000,
    "judgeModel": "gpt-4o-mini"
  }
}

Or via createWunderland():

import { createWunderland } from 'wunderland';

const app = await createWunderland({
  llm: { providerId: 'openai' },
  emergent: true,
  emergentConfig: {
    maxSessionTools: 10,
    sandboxTimeoutMs: 5000,
    judgeModel: 'gpt-4o-mini',
  },
});

When enabled, the agent gains access to the forge_tool meta-tool during conversation.

How It Works

Two Creation Modes

Compose mode chains existing tools into pipelines. No sandbox needed — only invokes pre-approved tools.

{
  "name": "research_and_summarize",
  "implementation": {
    "mode": "compose",
    "steps": [
      { "tool": "web_search", "input": { "query": "$input.topic" } },
      { "tool": "generate_text", "input": { "prompt": "Summarize: $prev.output" } }
    ]
  }
}

Sandbox mode writes novel JavaScript in an isolated VM (128MB memory, 5s timeout). Blocked APIs: eval, Function, require, process, child_process, fs.write*.

{
  "name": "parse_csv",
  "implementation": {
    "mode": "sandbox",
    "code": "function execute(input) { return input.csv.split('\\n').map(r => r.split(',')); }",
    "allowlist": []
  }
}

LLM-as-Judge Verification

Every forged tool undergoes LLM-as-judge review before activation:

Mode	When	What it checks
Creation review	First forge	Code safety, test correctness, determinism
Reuse validation	Each invocation	Output matches declared schema
Promotion panel	Tier upgrade	Independent safety + correctness reviewers

Tiered Promotion

Tier	Scope	Promotion Rule
Session	Current conversation only	Auto on creation + judge approval
Agent	Persisted across sessions	5+ uses, confidence > 0.8, panel approved
Shared	All agents	Human approval required (HITL gate)

Example: Agent Forges a Tool During Chat

User: I need to convert temperatures between Celsius and Fahrenheit frequently.

Agent: I don't have a temperature conversion tool, but I can create one.
       [Calling forge_tool...]

       I've created "convert_temperature" — a sandboxed tool that converts
       between Celsius, Fahrenheit, and Kelvin. It passed safety review.

User: Convert 100°F to Celsius.

Agent: [Calling convert_temperature...]
       100°F = 37.78°C

The tool starts at session tier. After 5+ successful uses with >0.8 confidence, the agent can automatically promote it to agent tier for persistence.

CLI Commands

wunderland emergent list              # List all emergent tools for the current agent
wunderland emergent inspect <id>      # Show source code, judge verdicts, usage stats
wunderland emergent export <id>       # Write a portable YAML/JSON package for reuse
wunderland emergent import <file>     # Import a portable package into another agent
wunderland emergent promote <id>      # Promote to shared tier (requires human approval)
wunderland emergent demote <id>       # Deactivate a tool (preserved for audit)
wunderland emergent audit <id>        # Show full audit trail

Export and Reuse

Emergent tools can be exported as portable agentos.emergent-tool.v1 YAML or JSON packages and imported into another agent.

• compose tools are portable by default
• sandbox tools are portable only when the package includes source code
• redacted sandbox exports remain useful for audit and Git review, but they are intentionally not importable into another runtime

Typical flow:

wunderland emergent export <id> --seed <seedId> --output ./my-tool.emergent-tool.yaml
wunderland emergent import ./my-tool.emergent-tool.yaml --seed <otherSeedId>

Safety Invariants

• Emergent tools cannot modify the guardrail pipeline
• Emergent tools cannot access other agents' memory or credentials
• Sandbox runs in an isolated V8 context with hard memory and timeout limits
• Forge decisions and metadata are logged to the audit trail
• Human approval is required for shared-tier promotion
• Raw sandbox source is redacted at rest by default unless you explicitly enable source persistence

Configuration Reference

{
  emergent: true,
  emergentConfig: {
    maxSessionTools: 10,       // max tools per session
    maxAgentTools: 50,         // max persisted per agent
    sandboxMemoryMB: 128,     // VM memory limit
    sandboxTimeoutMs: 5000,   // VM execution timeout
    promotionThreshold: { uses: 5, confidence: 0.8 },
    judgeModel: 'gpt-4o-mini',
    promotionJudgeModel: 'gpt-4o',
    allowedSandboxAPIs: [],   // e.g. ['fetch', 'crypto']
    fetchDomainAllowlist: [], // e.g. ['api.example.com']
  },
}