Software Engineering → AI Systems
Documenting what I learn while building AI systems.
I'm Sheva, a Staff Engineer exploring agent systems, enterprise AI adoption, harness engineering, and the challenges of turning foundation models into reliable business systems.
Former frontend engineer and engineering lead. Currently exploring how AI changes software, organizations, and the way we build systems.
Current Focus
These are the themes I am actively researching, experimenting with, and organizing notes around while building practical AI systems.
Why This Site Exists
Most of the content on this site originates from research sessions, implementation work, architecture reviews, and experiments conducted while building AI systems. Rather than presenting final answers, I document investigations, evolving mental models, and lessons learned along the way.
Explore Topics
Notes organized around the AI systems questions I am working through.
Agent Systems
Notes on agent loops, memory, tool use, orchestration, session behavior, and OpenClaw architecture.
Enterprise AI
Practical notes on AI platform architecture, governance, knowledge retrieval, runtime capabilities, and enterprise adoption constraints.
Engineering Leadership
Reflections on AI adoption, organizational design, harness engineering, technical leadership, and how engineering work changes around AI.
Recent Notes
Recent articles from active AI research and implementation work.
OpenClaw Concurrency Is Not Isolation
My investigation into why concurrent OpenClaw sessions can still compete for shared runtime resources.
Skill-Defined Tools vs Runtime-Provided Tools in Enterprise Agent Platforms
A platform architecture note on why skills can describe tool usage, but runtimes must own executable capabilities.
Building SOP Retrieval for OpenClaw Skills Instead of Prompt-Stuffing
A design note on separating OpenClaw skill workflow from SOP retrieval, search quality, and business policy ownership.
OpenClaw Permissions: Why Prompt Rules Are Not Enforcement
An investigation into why prompts guide behavior but runtime approval systems are needed for real enforcement.
Agent Loops vs Agent Teams: Understanding Where Intelligence Actually Lives
A practical distinction between agent loops, agent teams, execution style, topology, and where orchestration should own control flow.
Investigating Honcho Memory Architecture for OpenClaw
Notes on Honcho representations, observe_me, memory quality, and how long-term memory should be applied in OpenClaw.