Journal
AI
5 posts
- Building a Personal AI OS: File Watchers, Local IPC and Daemonised AgentsA terminal agent is a tool you visit. An AI OS is a set of small daemons that react to things happening on your machine. The difference is not the model — it is the event bus, the queue, and the discipline to keep every piece boring.AutomationAI8 min
- Autonomous CI/CD: Letting an Agent Triage the Build Without Letting It ShipMobile release engineering is mostly the same six failures on repeat — a provisioning profile, a Gradle version, a flaky simulator. That is exactly the shape of work an agent is good at, provided you are ruthless about which half of the pipeline it can touch.AutomationCI/CD7 min
- Context Window Engineering: Compaction, Retrieval and Why Long Sessions DegradeA million-token window did not solve the problem, it moved it. Attention is not uniform, irrelevant context actively costs accuracy, and the skill that matters now is deciding what an agent is allowed to see.ArchitectureAI8 min
- Hardening Agent Sandboxes: MicroVMs, gVisor and WebAssembly for Untrusted CodeAn AI agent running shell commands is an arbitrary code execution primitive you installed on purpose. Here is the isolation ladder — container, gVisor, Firecracker, WebAssembly — what each one actually stops, and what it costs per sandbox.AIAgents7 min
- Deterministic Multi-Agent Orchestration: Escaping the Infinite ReAct LoopAutonomous coding agents fail in a predictable way: context poisoning inside a freeform reason-act loop. The fix is not a better prompt — it is a finite state machine with programmatic gates, scoped tool access per phase, and automatic rollback.AutomationArchitecture9 min