Harness engineering: Leveraging Codex in an agent-first world — Harness builds a million-line product using only Codex, redefining engineers as prompt designers and environment builders. This proves agent-first production is viable and forces outcome engineers to treat prompt and environment engineering as core platform work (Principles 01, 03, 06).

Perplexity’s “Search as Code” lets AI models write their own search pipelines instead of calling fixed APIs — Perplexity lets models generate Python search pipelines inside a sandbox, improving accuracy and cutting token costs up to 85%. Outcome engineers should rethink retrieval architecture: let agents produce retrieval code to lower token spend and move retrieval into reproducible artifacts (Principles 06, 07).

Her · हेर — a detective for your Claude Code sessions — Her analyzes Claude Code session traces, reconstructs agent behavior, flags risky actions, and explains findings deterministically inside your private runtime. Run session forensics as part of your immune system to restore observability, attribute failures, and limit risky autonomous behavior (Principles 06, 14, 07).

Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering — The paper empirically shows agentic software engineering spends most tokens on iterative code review and input tokens, shifting optimization from generation to verification. Engineers must instrument token flows, optimize verification loops, and bake token-aware validation into cost and outcome audits (Principles 02, 16, 06).

Kyushu — A self-hostable WASM sandbox for JavaScript workers — Kyushu compiles JavaScript/TypeScript handlers into a single WebAssembly binary so you can run Cloudflare Workers–style runtimes on any VPS without Node or Docker. Use small, self-hosted WASM islands to sandbox agent-executed artifacts and simplify deployment and governance of agent-hosted workers (Principles 07, 08).