Latest

A ¥67 OpenCode Go trial exposed the gap between low API prices and affordable coding subscriptions. Our iKanban task comparison, the limits of API-value ratios, and why more tokens do not automatically become our own ability.

Borrowing the L1–L5 scale from self-driving cars, we ignore implementation details and look only at outcomes and the division of labor between humans and AI. The result is five levels, AL1 to AL5: code completion, conversational Q&A, a collaborator that explores on its own, a leader that proposes solutions beyond expert level, and a fully autonomous system with no human involved.

Why can the same GPT model stop midway, slow down, or struggle to finish a task in a different environment? Drawing on our experience with OpenCode and Codex, we examine how the harness affects availability and task performance, and how to choose a better fit.

Long-running AI agents can produce more code and research than we can absorb. That output extends our capacity, but it does not automatically become our knowledge, judgment, or ability.

A practical fuzzy-finding workflow that turns shell history, editor search, and browser history into layered indexes for quickly recovering commands, project context, and web resources.

Our first experience with DeepSeek Harness, why its everything-is-a-plugin architecture led us to rewrite iKanban on top of it, what composition made possible, and which missing workflow features still required our own solutions.

A Claude account ban forced our fallback stack back into the main role. This is how GPT, OpenCode, iKanban, and a skill-based AI Workflow kept work moving, where they still fall behind dynamic workflows, and why orchestration should become a portable plugin rather than a runtime feature.