Blog

Not all agent behaviour needs to be agentic. Agents that aggressively crystallize repeated tasks into deterministic workflows deliver better security, permissioning ergonomics, predictability, and token efficiency.

Long-running personal agents degrade in quality, safety, and predictability as context grows — and the standard fix (compaction) introduces its own failure modes that attackers can exploit.

Personal agents need private data, untrusted content, and external communication to be useful — the exact three conditions that make them indefensible. Every proposed fix costs you the thing you built the agent for.

Pre-PMF, learning velocity is the only KPI that matters. Everything else is either serving that goal or sabotaging it.

With AI eliminating the engineering throughput bottleneck, the new constraint is clarity—the PM's ability to aim the firehose through testable hypotheses and coherent product vision.

Most founders say "MVP" when they mean prototype. This confusion leads them to outsource work that cannot be outsourced—the iterative, judgment-heavy process of finding product-market fit.