Harness engineering: Leveraging Codex in an agent-first world. Harness reports building a million-line product using only Codex, recasting engineers as prompt designers and environment builders. This is a concrete agent-first shipping pattern for production systems — adopt its CI, context-engineering, and orchestration practices to scale agent teams (Principle 03).

Perplexity’s “Search as Code” lets AI models write their own search pipelines instead of calling fixed APIs. Perplexity runs models that generate Python search pipelines inside a sandbox, boosting accuracy and cutting token costs up to 85%. Treat retrieval as executable artifacts rather than static APIs — this shifts how you design retrieval, tooling, and safety boundaries (Principles 06 & 07).

Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering. The paper empirically shows agentic software engineering spends the bulk of tokens on iterative code review and input tokens, not initial generation. Reprioritize optimizations toward verification, input-efficiency, and review tooling to reduce operating cost and improve auditability (Principles 02, 06, 16).

When Claude changed, everything changed: Managing AI blast radius in production. A Claude Sonnet 4.5 upgrade silently broke API contracts, exposing gaps in gating, testing, and human-in-the-loop safeguards. Use this as a production playbook: enforce upgrade gates, contract tests, and rollback paths to contain model blast radius (Principles 10, 14, 15).

Her · हेर — a detective for your Claude Code sessions. Her reconstructs Claude Code session traces, flags risky actions, and explains agent behavior deterministically inside your private runtime. Add session forensics and deterministic explanation to your observability stack to investigate failures and prove what agents actually did (Principles 06, 07, 14).