Research
Evidence, not adjectives
Eight systems we have actually built, split by status: running in production, operationally hardened, or validated in research. Claims below are specific enough to check.
Systems
The stack
| System | Status | What it does |
|---|---|---|
| Multi-Agent Research Pipeline | Production | Four agents — Researcher → Analyst → Verifier → Writer. Produces 150+ line reports with market data, competitive intelligence and citations. 30+ sources per report, ~45 min delivery. |
| x402 Payment Protocol | Production | Pay-per-call API accepting USDC on Base. No KYC, no signup. 6 paid endpoints live plus a free tier at 5 calls/min. |
| Security Audit Framework | Operational | Automated scanning of agent pipelines for secret management, key exposure and injection surface. 11 vulnerabilities found on internal systems, 3 critical, all remediated. 7.5/10 internal score. |
| Agent Reputation System | Operational | Probabilistic reputation across 13 agent profiles. 500+ tasks completed at a 100% success rate, with reputation-weighted task prioritisation in production. |
| GoA — Graph of Agents | Research | Replication of Yun et al. 2026 (ICLR 2026). Selective agent routing: 3 agents chosen from a pool of 6 outperform all 6. Validated on MMLU, MMLU-Pro, GPQA, MATH, HumanEval and MedMCQA. |
| SPIRAL RL + D-GAT Coordination | Research | Self-play reinforcement learning for multi-agent coordination, with distributed coordination via D-GAT (DG-MAPPO). Proven in simulation, pending production integration. |
| Cognitive Architecture (FEP) | Research | Free Energy Principle implementation: a world model of 5,600 entities, self-model, ICL cycles and UPAL learning. Code complete, not yet in product use. |
| Probabilistic Market Forecasting | Research | Thompson sampling plus probabilistic model averaging for Polymarket paper trading. 157 snapshots and 64 trade entries analysed, with P(profit) per position. Paper trading only. |
Writing
Posts and benchmarks
Published analysis of agent security, model benchmarks and trust in high-stakes decisions.
Benchmark
GLM 5.2 Security Benchmark
72% of AI agents fail basic secret management checks. Analysis across 200+ production agents.
Read → SecurityWhy AI agents need security audits
A Codex sensitive-file exclusion issue, and GLM beating Claude on security benchmarks.
Read → TrustAI agents now read MRI scans
A Claude Code analysis hit #6 on Hacker News. What it means for high-stakes decisions.
Read →