Claude Code Vulnerability Could Let Attackers Steal Credentials From GitHub, Says Microsoft — Microsoft discloses a prompt-injection flaw in Anthropic’s Claude Code GitHub Action that could expose CI/CD credentials before Anthropic released a patch. Outcome engineers must treat agent actions inside CI as a high-risk attack surface: secrets handling, sandboxing, and an immune-system posture are non-negotiable (Principles 14, 15).

When Claude changed, everything changed: Managing AI blast radius in production — A Claude Sonnet model upgrade silently broke API contracts, revealing gaps in gating, testing, and human-in-the-loop safeguards. This underscores that model upgrades are infra changes — add contract tests, canaries, and explicit human gates to limit blast radius (Principles 10, 14, 15).

Harness engineering: Leveraging Codex in an agent-first world — Harness reports shipping a million-line product using only Codex, recasting engineers as prompt designers and environment builders. Use this as a playbook: invest in context engineering, developer-facing islands, and agent-friendly interfaces so agents become reliable delivery lanes (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 sandboxed Python search pipelines, boosting retrieval accuracy and cutting token costs by as much as 85%. Outcome engineers should adopt programmable, sandboxed retrieval as a first-class primitive to reduce tokens, improve relevance, and contain risky tool calls (Principles 06, 07).

Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering — Empirical analysis shows agentic software engineering spends most tokens on iterative code review and input/context, not initial generation. That shifts optimization: prioritize verification pipelines, caching, and cheaper checks over brighter-but-costlier generation strategies when designing agentic systems (Principles 02, 06, 16).