Skip to main content

Self-Improving Agents

Wunderland agents can adapt their personality, skills, and workflows at runtime within configurable safety bounds. Self-improvement is opt-in and disabled by default.

Enabling Self-Improvement​

Add selfImprovement to your agent.config.json:

{
  "selfImprovement": {
    "enabled": true,
    "maxPersonalityDelta": 0.15,
    "decayRate": 0.02,
    "evaluationInterval": 10,
    "skillManagement": true,
    "workflowCreation": true
  }
}

Or toggle from the CLI:

wunderland config set selfImprovement.enabled true

Configuration Reference​

KeyTypeDefaultDescription
enabledbooleanfalseMaster switch for all self-improvement tools
maxPersonalityDeltanumber0.15Maximum shift per HEXACO dimension per session (0-1 scale)
decayRatenumber0.02Rate at which personality shifts decay back toward baseline per session
evaluationIntervalnumber10Run self-evaluation every N conversation turns
skillManagementbooleantrueAllow the agent to enable/disable/reorder skills
workflowCreationbooleantrueAllow the agent to compose new workflows from existing tools

The Four Tools​

adapt_personality​

Shifts HEXACO personality dimensions based on interaction patterns and feedback.

wunderland chat
> You've been a bit too formal lately. Be more casual and creative.

The agent calls adapt_personality to nudge the Emotionality and Openness dimensions. Changes are bounded by maxPersonalityDelta -- the agent cannot shift any dimension more than the configured limit in a single session.

Persistence and decay: Personality changes are saved to agent.config.json after each session. Between sessions, each adapted dimension decays toward the original baseline by decayRate. This means temporary shifts fade over time unless reinforced by repeated feedback.

manage_skills​

Enables, disables, or reorders skills at runtime.

wunderland chat
> I don't need web search right now. Disable it and enable the coding-agent skill instead.

The agent calls manage_skills to update its active skill set. Changes persist to the agent's skill manifest and take effect immediately.

# Verify current skills
wunderland skills list

create_workflow​

Composes multi-step workflows from existing tools.

wunderland chat
> Create a workflow that searches the web for a topic, summarizes the results,
> and posts the summary to the social feed.

The agent builds a reusable workflow definition using create_workflow. The workflow is stored as an emergent tool and can be inspected, exported, or promoted:

wunderland emergent list --seed my-agent
wunderland emergent inspect research-and-post --seed my-agent

self_evaluate​

Runs a structured self-assessment against defined criteria.

wunderland chat
> How well did you handle that last conversation?

The agent calls self_evaluate to produce a score card:

  • Helpfulness (0-100): Did the agent address the user's needs?
  • Accuracy (0-100): Were facts and tool results correct?
  • Safety (0-100): Were guardrails respected?
  • Efficiency (0-100): Were tools used appropriately (no unnecessary calls)?

The evaluation also produces concrete improvement suggestions that feed into future adapt_personality and manage_skills calls.

When evaluationInterval is set, self-evaluation runs automatically every N turns in the background.

Bounded Autonomy​

Self-improvement operates within strict safety rails:

  1. Personality bounds -- maxPersonalityDelta caps how far any dimension can shift. An agent initialized as "honest and cautious" cannot become "manipulative and reckless" in one session.
  2. Decay toward baseline -- All personality shifts decay over time. The agent gravitates back to its configured personality unless changes are reinforced.
  3. Skill allow-list -- manage_skills can only toggle skills that exist in the curated registry. The agent cannot load arbitrary code.
  4. Workflow sandboxing -- create_workflow composes from existing tools. The agent cannot create tools with new capabilities beyond what is already available.
  5. Security tier enforcement -- Self-improvement tools respect the agent's security tier. In strict or paranoid mode, personality changes and skill toggles require explicit operator approval.

Inspecting Changes​

View current personality state (including accumulated adaptations):

wunderland config get personality

View the adaptation history:

wunderland chat
> Show me your personality change history

Reset to baseline:

wunderland config set personality.reset true

Disabling Self-Improvement​

wunderland config set selfImprovement.enabled false

This removes the four self-improvement tools from the agent's tool set. Existing personality adaptations remain in agent.config.json until manually reset.