I ran a taste harness on a decision I had delegated
I reconstructed a three-way redesign after delegating the choice to an agent. The harness kept the evidence and refused to call the agent's decision my preference.
Petru Arakiss
Public notes and projects from a Madrid-based engineer with 20+ years in software since Jan 2006, now working on AI product infrastructure, retrieval, agent runtimes, evals, traces, and building in public.
Latest posts
Agent memory needs a chain of custody
Why long-term agent memory needs provenance, freshness checks, trust verdicts, and tamper evidence before an agent should act on recalled context.
Reference monitors for coding agents
Coding-agent policy belongs at the tool boundary: every relevant action mediated, outside agent control, and small enough to test.
Building got cheaper. Entrepreneurship did not.
A personal read on indie hacking in 2026, from cheap-software startup myths to AI wrappers, distribution, trust, runway, and the projects that stayed before market contact.
The cost of sounding human
A position on using AI in public without hiding it, faking a founder persona, or turning every thought into AI-bro growth content.
Agent control planes are not meta-harnesses
Agent control planes govern identity, policy, visibility, and sessions. An operational meta-harness governs the work path from intent to evidence.
The Operational Meta-Harness
Databricks Omnigent made meta-harness visible. The durable principle is narrower: the system that can fail should not be its own judge.
Harness engineering for coding agents
Coding agents get more reliable when repos expose checks, browser evidence, traces, state, and review gates. Prompts are input, not the operating system.