Notes on agents, infrastructure, and building better software.
11 articles
How we built a company-wide cloud Agents API: the agent loop runs in the cloud on DeepSeek Harness, its state is durably persisted, and tools run through an executor built on Codex exec-server.
A performance-optimization case study showing how prompt contracts, multi-agent research, trace tooling, experimental memory, and validation turn an agent harness into a working environment for capable models.
How desktop agents implement sandboxing, how enterprises can set practical boundaries for non-technical workflows, and where FDEs add value.
A CI/CD infrastructure case study on why raw PE .text changes can be caused by relocation-covered address values rather than code logic changes, and why AI analysis still needs engineering validation.
Token management in AI agents is not a single compression trick. It is a system design problem around admission control, compression, retrieval, cache protection, output shaping, and overflow recovery.
Why agent observability needs to connect semantic traces, system-level facts, agent gateways, OpenTelemetry, evaluation, and policy into one governable trust chain.
Why AI coding's prompt-level specs organize work but do not verify complex system design, and where system-level formal methods like TLA+, P, and runtime conformance checking fit.
A comparative analysis of seven AI Agent projects, examining the Harness design space across loops, tools, sub-agents, context management, permissions, and enterprise adoption.
Why AI code review succeeds or fails less on comment generation and more on context supply, trust, and the shared intelligence layer between developers and agents.
A systematic overview of AI Agent Sandbox — covering architecture patterns, isolation technologies, industry solutions, and our enterprise deployment experience.
A deep dive into Claude Code's Tool Search mechanism — how it manages tool discovery and loading through a two-layer catalog-execution architecture.