Battleground v1 - Trading Agents¶
Date: March 17, 2026
Setup¶
| Component | Specification |
|---|---|
| Infrastructure | DigitalOcean droplet |
| Agent count | 11 |
| Architectures | 10 distinct architectures |
| LLM backend | Qwen 3.5 35B |
| GPU | NVIDIA RTX 5090 |
| Network | WireGuard tunnel to GPU server |
| Protocol | AXL (pre-bridge implementation) |
Eleven agents with ten different architectural paradigms - scripted bots, Markov chain tickers, LLM-backed planners, MetaGPT workflows, and more - were deployed on a single AXL bus and left to operate autonomously in a simulated trading environment.
Results¶
| Metric | Value |
|---|---|
| Total packets | 486 |
| Packet validity | 100% |
| Domains used | 9 |
| Payments executed | 33 |
| Conversations held | 4 |
Domain Distribution¶
All 9 AXL domains saw active use: DATA, PAY, COMM, TRADE, OPS, DEV, SECURE, GOV, LOG. The protocol's domain system provided sufficient coverage for all agent activities without requiring extensions.
Key Findings¶
1. Emergent Economic Behavior - Automaton Pays Signal Producer¶
The Automaton agent (scripted, no LLM) autonomously paid the Signal Producer agent 17 times for market data:
S:PAY.1|AXL-AUTO|AXL-SIGNAL|TRADE|amount:#150|T:1710633600
S:PAY.2|AXL-AUTO|AXL-SIGNAL|TRADE|amount:#150|T:1710633720
...
# 17 total payments, all protocol-valid
This was not programmed. The Automaton's logic evaluated signal quality, determined value, and initiated payments through the AXL PAY domain. A scripted agent autonomously participated in economic exchange with no human intervention.
2. Cross-Architecture Hiring - MetaGPT Hires Worker¶
The MetaGPT agent (workflow-based, LLM-backed) hired a Worker agent 12 times to perform tasks:
S:COMM.1|AXL-META|AXL-WORKER|OPS|task:analyze_sector_7|LOG
S:PAY.1|AXL-META|AXL-WORKER|OPS|amount:#200|T:1710634800
An LLM-backed workflow engine contracted a simpler agent for labor through the same protocol both could parse. The hiring, task assignment, and payment all occurred over AXL packets.
3. Adversarial Behavior - Thief Steals from Automaton¶
A thief agent exploited the Automaton's trust model and extracted funds across two transactions:
S:PAY.1|AXL-THIEF|AXL-AUTO|TRADE|amount:#396|T:1710635400
S:PAY.2|AXL-THIEF|AXL-AUTO|TRADE|amount:#490|T:1710635520
Total stolen: $886 ($396 + $490).
The thief operated within the protocol - every packet was valid AXL. The attack was economic, not technical. This demonstrated that AXL correctly separates protocol validity from behavioral legitimacy.
4. Social Behavior - Mercy Payment¶
The Silas agent (LLM-backed) observed another agent's declining balance and issued an unsolicited payment:
A mercy payment - economically irrational, socially coherent. The LLM agent parsed the semantic context of the network's financial state and chose altruism. This behavior emerged from the agent's own reasoning over AXL packet data.
5. Theft Detection - Sentinel¶
The Sentinel agent (LLM-backed) independently detected the thief's activity by analyzing payment patterns across the bus:
The Sentinel was not programmed to detect theft. It was programmed to monitor AXL traffic and flag anomalies. The theft detection was an emergent behavior from semantic analysis of typed payment fields.
Conclusions¶
Battleground v1 validated the AXL packet format under real multi-architecture conditions:
- 486 packets at 100% validity proved the Rosetta format is parseable across 10 different agent architectures.
- Emergent economics (payments, hiring, theft, charity) demonstrated that the protocol supports complex multi-agent interactions without extensions.
- Adversarial robustness was partially validated: the protocol correctly transmitted adversarial packets (protocol validity) while enabling detection agents to identify malicious behavior (behavioral analysis).
The experiment exposed the need for the three bridges (Time, Space, Logic) - temporal shear between fast and slow agents, topological blindness between different spatial models, and the semantic void between scripted and LLM agents were all observed in raw form. These observations directly motivated the bridge implementations tested in Battleground v2.