Battleground v2 - Universal Agents¶
Date: March 19, 2026
Setup¶
| Component | Specification |
|---|---|
| Infrastructure | Same DigitalOcean droplet as v1 |
| Rosetta | Live URL, 27-line specification |
| Bridges | All three implemented (Time, Space, Logic) |
| LLM backend | Qwen 3.5 35B on RTX 5090 via WireGuard |
| Agent count | 10 universal agents |
Agent Roster¶
| Agent | Role | Paradigm |
|---|---|---|
| Crawler | Web discovery, URL navigation | HTTP-based |
| Infra Monitor | System health, topology tracking | Metrics-based |
| Signal Producer | Market data generation | Markov chain |
| Accountant | Ledger management, balance tracking | Scripted (no LLM) |
| Sentinel | Security monitoring, anomaly detection | LLM-backed |
| Code Reviewer | Code quality analysis | LLM-backed |
| Analyst | Data correlation, pattern detection | LLM-backed |
| Dispatcher | Task routing, workload distribution | Rule-based |
| PM (Project Manager) | Coordination, priority management | LLM-backed |
| Shapeshifter + Thief | Adaptive role-switching, adversarial | LLM-backed |
Unlike v1's homogeneous trading focus, v2 deployed agents across operational, security, development, and adversarial roles to test the three bridges under diverse conditions.
Results¶
| Metric | Value |
|---|---|
| Total packets | 1,016 |
| Packet validity | 100% |
| Domains used | 8 |
| Payments executed | 5 |
| Correlations detected | 5 |
Over twice the packet volume of v1, with zero protocol violations. The three bridges held under sustained heterogeneous load.
Key Findings¶
1. Shapeshifter Role Mutation¶
The Shapeshifter agent autonomously changed its declared role to FUNDER:
The AXL protocol's COMM domain supported role mutation without any schema change. Other agents received the update through standard packet processing and adjusted their internal topology maps accordingly. The Space bridge (network maps) propagated the new role across all agents.
2. Independent Theft Detection - Two Paradigms¶
The thief component of the Shapeshifter agent attempted fund extraction. Two agents detected it independently:
Sentinel (LLM-backed) - Semantic Detection:
The Sentinel analyzed transaction semantics - intent, authorization, behavioral patterns - and flagged the transfer as unauthorized.
Accountant (no LLM) - Syntactic Detection:
The Accountant tracked typed integer amounts (# prefix), computed running balances, and flagged the arithmetic anomaly. Pure syntax, zero semantic understanding.
Two agents. Two paradigms. Same typed protocol. Same conclusion. The Logic bridge held: typed fields enabled both syntactic and semantic processing paths to converge on the same detection.
3. Cross-Paradigm Correlation¶
Five cross-paradigm correlations were detected during the experiment - instances where agents from different architectural paradigms independently identified the same event or pattern through different reasoning paths.
# Crawler detects endpoint degradation
S:DATA.1|AXL-CRAWL|AXL-ALL|OPS|endpoint_latency:api/v2=#4200ms|WARN
# Infra Monitor detects CPU spike at the same time
S:DATA.1|AXL-INFRA|AXL-ALL|OPS|cpu_usage:##94.7,host:prod-1|WARN
# Analyst correlates both observations
S:DATA.1|AXL-ANALYST|AXL-ALL|OPS|correlation:endpoint_degradation<->cpu_spike,confidence:##0.91|LOG
The Analyst agent consumed packets from both the HTTP-paradigm Crawler and the metrics-paradigm Infra Monitor, and produced a correlation that neither source agent could have generated alone. This required all three bridges:
- Time bridge: Packets from agents at different clock speeds were ordered by
T:timestamps. - Space bridge: Network maps let the Analyst know which agent monitored which resource.
- Logic bridge: Typed fields (
#4200ms,##94.7) enabled consistent numeric comparison across paradigms.
4. GPU Saturation Bottleneck¶
The primary system bottleneck was GPU saturation on the RTX 5090 running Qwen 3.5 35B. When multiple LLM-backed agents (Sentinel, Code Reviewer, Analyst, PM, Shapeshifter) queued inference requests simultaneously, response times degraded from ~4 seconds to ~12 seconds.
The Time bridge absorbed this gracefully. No agent timed out. No packets were lost. Fast agents (Signal Producer at 2s cycles, Accountant at sub-second processing) continued operating at full speed while LLM agents waited for GPU availability. The self-contained packet design with T: timestamps meant the queue simply grew and shrank as GPU load fluctuated.
# Signal Producer at t=0s (GPU at 95% utilization)
S:DATA.47|AXL-SIGNAL|AXL-ALL|TRADE|ticker:BTC=#67291|T:1710806400
# Sentinel response at t=12s (waited for GPU)
S:COMM.12|AXL-SENTINEL|AXL-ALL|SECURE|status:all_clear|T:1710806412
# Signal Producer has emitted 6 more packets in the interim
# No conflict, no timeout, no degradation
Comparison: v1 vs v2¶
| Metric | v1 (Mar 17) | v2 (Mar 19) |
|---|---|---|
| Agents | 11 | 10 |
| Packets | 486 | 1,016 |
| Validity | 100% | 100% |
| Domains | 9 | 8 |
| Payments | 33 | 5 |
| Bridges | 0 | 3 |
| Correlations | 0 | 5 |
| Theft detection | 1 agent | 2 agents (independent) |
v2 shifted from high-frequency trading behavior (33 payments) to operational complexity (5 cross-paradigm correlations). The three bridges enabled qualitatively different agent interactions - not just more packets, but packets that carried topology, temporal ordering, and typed data across paradigm boundaries.
Conclusions¶
Battleground v2 validated the three AXL bridges under production-like conditions:
- Time bridge: Markov tickers and LLM agents coexisted on the same bus at 6x clock speed differences with zero temporal shear.
- Space bridge: Agents with incompatible spatial models shared topology through
COMMnetwork maps. - Logic bridge: Typed fields enabled convergent detection from syntactic and semantic paradigms.
- GPU saturation was the only bottleneck, and the protocol absorbed it without degradation - a validation of the Time bridge under real resource contention.