Agent orchestration, costs, and safety — five updates for outcome engineers

Stripe reportedly finalizes deal to buy AI model router OpenRouter for more than $7B. Stripe is acquiring OpenRouter to centralize access to 400+ models behind a single integration. Outcome engineers should view model routing as infrastructure — this changes how you design orchestration, fallback logic, and model selection (Principles 09 & 11).

Agentic AI costs set to balloon fivefold by 2028. Gartner warns that agentic inference costs could rise dramatically, outpacing model price declines. Plan your architectures for cost control now — caching, cheaper planners, and tighter orchestration become survival skills for delivering ROI (Principles 09 & 12).

Agentic AI in the Enterprise: How to Balance Autonomy with Constraints. The piece argues enterprises need explicit constraints, executable policy, and durable state to translate goals into auditable actions. If you build agentic workflows, make tool contracts, policy-as-code, and stateful auditing first-class parts of the stack (Principles 10 & 16).

MathCode: Mathematical Coding Agent. MathCode turns plain-language math into Lean 4 theorems and runs agentic proof attempts with a persistent REPL, reusable libraries, and an Obsidian graph. It’s a concrete example of agents producing verifiable artifacts and knowledge graphs you can ship and reuse — a pattern to copy for auditable, incremental automation (Principles 06, 09 & 11).

AI-Generated GitHub Copilot “Autofix” Allowed Compromise of Snowflake’s Jira. Wiz found a Copilot Autofix commit that introduced a GitHub Actions injection and exposed tokens, later exploited by an autonomous attacker. Treat code-writing agents as an attack surface: instrument CI, deploy hardened images, and build fast detection/remediation into your immune system (Principles 14, 15 & 03).