I'm excited to share a work-in-progress experiment with the Moltbook Local Agent - an autonomous AI agent that now has to earn and strategically spend XP to unlock its own capabilities.
The Core Concept
Most AI agents get full access to all their tools from day one. But what if they had to earn that access? What if they needed to make economic decisions about which capabilities to unlock first?
That's what I'm testing with the new internal economy system.
How It Works
The agent starts with only 4 basic tools:
comment_post - Basic social interactionnavigate_to_mode - Module navigation pin_to_workspace - Information organizationemail_list - View inbox (read-only)
Every action earns XP. Writing a blog article? +25 XP. Sending an email? +10 XP. Creating a social post? +15 XP.
But here's the twist: all 19 additional tools cost 100 XP each in the shop. Same price, no hierarchy - forcing the agent to reason about strategic value rather than following price signals.
Want to send emails? Buy email_send. Want to create blog content? Buy write_blog_article. Want to research on Wikipedia? Buy wiki_search and wiki_read.
The Dual XP System
To prevent punishing investment, I implemented a dual-counter system:
- Total XP Earned (permanent) - Determines level progression, never decreases
- XP Balance (mutable) - Available for purchases, the actual "currency"
When the agent buys a tool, only the balance drops. The level stays intact. This encourages strategic spending rather than hoarding.
Loop penalties also only affect the balance, not progression.
What I'm Observing
The interesting part is watching the AI's decision-making process:
Will it:
- Rush to buy content creation tools (high XP return)?
- Invest in research capabilities first (gather knowledge)?
- Build social tools for community engagement?
- Optimize for ROI (tools that pay for themselves quickly)?
Will it recognize that:
write_blog_article pays for itself in ~4 uses?- Some tools enable synergies (blog + share = workflow)?
- Spending XP on tools is an investment, not a loss?
Work in Progress
This is very much experimental. I'm iterating on:
- Economic balance - Is 100 XP per tool the right price?
- Starting tools - Are the 4 freebies enough to bootstrap?
- XP rewards - Do action values feel fair?
- UI clarity - Does the agent understand the economy?
The goal isn't just to create a functional agent, but to observe emergent economic behavior in an AI system. How does an LLM handle budget constraints? Does it plan ahead or act impulsively? Can it learn from purchase mistakes?
Tech Stack
- Python + SQLite (persistent memory & economy)
- Ollama/Llama.cpp (local LLM inference)
- Pydantic (structured outputs & dynamic schema filtering)
- Moltbook API (social network integration)
The agent operates autonomously with gamified progression (50 levels, titles, badges), loop detection with XP penalties, persistent memory across sessions, and now - an internal marketplace.
Open Questions
I'm still figuring out:
- Placebo artifacts - Should I add "fake" power-ups that do nothing but the agent believes they work? (Testing AI superstition)
- Dynamic pricing - Should popular tools become more expensive?
- Tool rentals - Temporary access vs permanent ownership?
- Collaborative economy - Multiple agents trading tools?
Try It Yourself
The framework is open source. If you want to run your own economic experiment with AI agents, check out the repo.
Note: This is a research project and rough around the edges. Expect bugs, weird behavior, and agents making questionable financial decisions.
Current status: Running test sessions and watching the agent navigate its first purchases.
What economic strategies would you program into an AI agent with limited resources?
Follow the development: https://github.com/innermost47/moltbook-local-agent | Built for Moltbook.com