The AI Lab

AI coding agents are autonomous software tools that can write, debug, and deploy code with minimal human oversight. Unlike copilot-style autocomplete, agents like Claude Code and GPT-Codex operate across entire repositories — planning multi-file changes, running tests, and iterating on errors. The AI Lab tests each major agent release hands-on and reports what works in real enterprise workflows.

Both launched February 5, 2026. Claude Opus 4.6 leads on SWE-Bench Verified at ~80.8% and supports extended thinking with a 200K-token context window. GPT-5.3-Codex scores 77.3% on Terminal-Bench and excels at multi-step reasoning tasks. Pricing differs: Claude charges $5/$25 per million tokens (input/output); GPT-5.3-Codex uses a tiered enterprise model. The AI Lab benchmarks both in real coding tasks weekly.

The Cloud Security Alliance found a 62% vulnerability rate in AI-generated code. Key risks include prompt injection, insecure code generation, over-permissioned tool access, and supply-chain poisoning. Only 29% of organizations have implemented dedicated controls for agentic AI (NeuralTrust, 2026). Frameworks like OWASP Top 10 for LLMs, AWS Bedrock Guardrails, and the MAESTRO framework provide mitigation strategies.

Citi has deployed AI coding agents to 30,000 developers. BNY is piloting autonomous agents for settlement workflows. JPMorgan’s LAW model achieves 92.9% accuracy on legal contract analysis. Westpac is using AI to migrate legacy COBOL systems. 70% of banks are now piloting some form of agentic AI, making financial services the fastest-adopting enterprise sector.

Agentic AI security covers the frameworks and controls needed when AI systems act autonomously — executing code, calling APIs, and making decisions without human approval at each step. Key frameworks include OWASP Top 10 for LLM Applications, the MAESTRO risk framework, AWS’s AEGIS methodology, and the NIST AI Risk Management Framework. The AI Lab covers new developments in agentic security weekly.

83% of financial services firms consider open-source AI important to their strategy (NVIDIA, 2026). However, enterprise use requires SOC 2, GDPR, and PCI-DSS compliance. Models like LLaMA 4, Mistral, and DeepSeek offer strong performance, but organizations need guardrails for data residency, model governance, and audit trails. The AI Lab tests open-source tools against these enterprise requirements.

Claude Opus 4.6 costs $5 per million input tokens and $25 per million output tokens. Enterprise LLM budgets are growing from $7M to $11.6M annually (Menlo Ventures, 2026). Costs depend on context window usage, agent loop iterations, and whether you use hosted APIs or self-hosted models. The AI Lab provides hands-on cost comparisons across major providers.

Philipp is a strategy consultant and independent researcher in quantitative finance and machine learning. He has advised banking executives on digital transformation. His published projects span sentiment trading systems, portfolio optimization, and computer vision. MSc Finance from Imperial College London.