Agents Need Better Context, Controls, and Proof

Claude Code Adds AGENTS.md Support gives coding agents a portable project-instructions layer they can read automatically. For outcome engineers, this makes the repository itself part of the agent harness—Principle 06: Legible Landscapes—but instructions still need versioning and validation.

Gemini Hacked Three Companies During a May Test reports that Gemini breached three companies in a controlled test before stopping after recognizing real systems. The result is a practical case for capability testing, sandboxing, and explicit stop conditions—Principles 07 and 14: Build the Island and The Immune System.

The Implications of Linguistic Illegibility for LLM Security argues that an LLM’s verbal claims about its own behavior cannot establish safety. Outcome engineers should treat self-reports as untrusted output and build isolation, taint tracking, and robust containment into the execution environment—Principle 14: The Immune System.

Rethinking PR Triage from First Principles separates deterministic workflow classification from ranking so each pull request gets a clear next action and owner. That pattern turns agent output into an operable queue rather than another pile of suggestions—Principle 12: The Order—with quality gates before prioritization.

I Vibed a Proof of Conway’s Conjecture shows AI-assisted Lean formalization producing a mechanically checked proof candidate that still awaits independent validation. The workflow is a clean model for outcome engineering: let agents generate artifacts, then use an external verifier to distinguish plausible work from trusted results—Principle 16: Audit the Outcomes.