Most AI agent frameworks today — the ones you’ve seen all over your feed — have a fundamental problem. If their results are within 10% of a hashmap, all they've really built is a slow, expensive wrapper around a hashmap.
They rely on sequential LLM calls, massive API bills, and brittle logic that falls apart the moment the environment gets messy.
I decided to fix that.
I built Ebbforge: A high-performance swarm intelligence engine written in Rust that runs 10 million agents locally with zero API cost.
Here are the 8 benchmarks that traditional architectures fail, but Ebbforge passes.
The Demo: 1,000 Agents, 60 FPS, Zero LLM Calls
We aren't just moving dots on a screen. Every agent in this swarm is a self-healing, learning unit powered by Temporal Difference Reinforcement Learning (TD-RL) and biologically-inspired memory decay.
The 8 Unsolved Problems
We ran Ebbforge against standard LLM-centric architectures on 8 tests that define real intelligence.
1. The Intelligence vs. Hashmap Challenge
The Goal: Catch an attacker who adds "padding" to a sequence to evade detection.
- Standard RAG/LLM: Misses the pattern (Critical False Negative).
- Ebbforge: Uses Longest Common Subsequence (LCS) math to recognize the structure of danger, even with noise. Blocked
2. The Groundhog Day Test
The Goal: Learn from a single failure and never repeat it.
- Most Agents: Loop and fail 9 times in a row.
- Ebbforge: One failure creates a persistent safety pattern. 9/9 subsequent attempts blocked
3. Cascade Failure Recovery
The Goal: Kill 30% of agents mid-flight and see if the swarm survives.
- Standard Systems: Corrupt their state or crash.
- Ebbforge: Survived 300 concurrent agent hard-kills and self-healed. 70% completion rate sustained
4. Organic Caste Emergence
The Goal: Can behavioral specialists emerge without hardcoded rules?
- Standard Systems: Require "Specialized Prompting."
- Ebbforge: Start with identical agents. After 500 ticks, they naturally split into "Brokers," "Hoarders," and "Neutrals" based on physics and reward pressure alone.
(See the other 4 benchmarks on the repo!)
Why Rust?
To handle 10 million agents, you can't be waiting for Python's Global Interpreter Lock (GIL) or $0.01-per-token API calls.
Ebbforge uses:
- AVX2 SIMD for physics.
- Rayon for grid-partitioned parallel processing.
- Zero-Copy Memory for agent communication.
At a Glance (The TL;DR)
| Challenge | Traditional Architectures | Ebbforge |
|---|---|---|
| Survive partial system failure | Failure: Corrupt state | Success: Self-heals |
| Learn from 1 failure | Failure: Repeats mistake | Success: 9/9 blocked |
| Traumatic memory retention | Failure: Equal decay | Success: 70,000x ratio |
| 10M agent coordination | Failure: O(N) flood | Success: Spatial wavefront |
Try It Yourself
The project is now live on GitHub as a pre-compiled binary. You can run the glassmorphism demo locally on any Linux x86_64 machine.
GitHub: juyterman1000/ebbforge-swarm-intelligence
I'm looking for feedback from the Rust and AI research community. If you've ever felt that agent frameworks are too slow or too "faked," Ebbforge is for you.
P.S. We just launched on Hacker News! Check out the discussion there too.
Top comments (0)