Why I forked localterm into TermDeck: a self-hosted browser workspace for agent-heavy terminal work. Live grid of sessions, clickable file paths, tweak mode that sends CSS selectors back to the agent, a markdown wiki, and a tldraw canvas.

Mastra shipped file-based agents, skills, schedules, and signals in nine days. My markdown vault has run all four DIY for months.

arXiv:2605.26112 says agent performance now comes from six system components around the foundation model: memory, context construction, skill routing, orchestration, and governance. I ran the paper through my personal agent setup as a self-audit. Verdict: half the paper was already implemented through months of small adoptions, two named failure modes (stale-but-confident memory, confident-but-unchecked skills) were real gaps, and both got fixed the same afternoon by the agent itself. This is the full process: paper to audit to shipped enhancements to this post.

I'm the lead product manager building Katalon True Platform, an AI-native testing platform, and every release my AI agent tests it alongside our QE team. One terminal command runs the full software testing loop: collect the release scope, design test cases, execute with the AI test runner and Playwright, file findings back. This is how a PM learns the product's real capabilities and limitations, with real screenshots, video, and a narrated deep dive for QA engineers, QE, and testers.

With GPT and Claude, any PM or designer can build something that technically works in half a day. It is addictive, and it lies. 'Technically works' (in isolation, on happy-path data, on my machine) is a different universe from 'works well inside the current system, at scale.' That second thing is the real engineering, and AI quietly lets you skip all of it. This post is about the prototype trap, the naivety of confusing a working demo with a production-ready feature, and the reframe that makes non-engineers actually useful with AI: run the loop to evaluate options and build conviction, spark the team instead of pretending the hack is done, and treat innovation as selection rather than invention.

A builder's argument for why neither product teams nor everyday users are ready for fully autonomous AI agents: verification is the expensive part and most of us can't afford it, and unverified AI output compounds like debt. Includes the snowball failure modes I hit running an agent-heavy personal knowledge system, and the practices I stole from gbrain and gstack to contain them: memory gates, provenance, supersede-not-append, periodic reindex, fresh-context isolation, and poisoning defenses.