Sakana AI launches Fugu, a multi-agent model API aimed at the export-control era. Sakana launches Fugu, a learned multi-agent orchestration API that manages model pools and routes work across agents as a hedge against export-control disruptions. Outcome engineers must treat model pools and learned routers as first-class infra — design for graceful degradation, fallbacks, and audit trails (Principle 09).

Researchers introduce Self-Harness, a framework that lets AI agents rewrite their own rules, boosting performance up to 60%. Self-Harness lets agents autonomously edit their operating harness using trace-driven edits and regression testing, improving performance by up to 60%. If you build harnesses, plan for safe self-modification: add CI-like regression suites, versioned harness artifacts, and pre-merge validation to avoid silent behavioral drift (Principle 16).

Temporary Cloudflare Accounts for AI agents. Cloudflare now offers 60-minute ephemeral Workers deployments without an account, creating frictionless sandboxes for agent experiments and demos. Use ephemeral, least-privilege runspaces to iterate faster and to contain blast radius during agent testing — a simple path to safer agent development workflows (Principle 07).

Stop wasting tokens and re-explaining your project between sessions. Recall stores fully-local, offline session summaries for Claude Code so projects resume without re-explaining context or burning tokens. Persistent, local project state is a practical lever for reliable multi-session agents: it reduces token costs, preserves privacy, and makes intent legible across handoffs (Principle 06).

Agentics: Cost to Implement vs Cost to Verify. The essay argues verification costs—not raw model throughput—determine when delegating work to agents is safe. That reframes your resourcing: invest in verifiers, monitoring, and human-in-the-loop checkpoints equal to agent development time to keep outcomes auditable and trustworthy (Principles 15 & 16).