Build agent workflows around proof, permissions, and judgment

ICE IT Shop Looks to Build an Agentic Software Factory separates agent-generated code from independent assurance and production authorization. That is Principles 09, 14, and 15 made concrete: coordination, immune defenses, and a human-controlled gate belong in the architecture.

How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering? asks whether autonomous ML work needs elaborate orchestration or whether a capable coding agent with direct execution access is enough. Test the simplest harness against real outcomes before adding layers of coordination—Principle 09.

Introducing Clef: Open-Source Decision Models and an RL Fine-Tuning Platform introduces fast, typed decision models that can route uncertain choices to people. That gives outcome engineers a practical pattern for keeping agents decisive on bounded tasks without making every decision autonomous—Principle 15.

SCLATE: A Substrate for Continual-Learning Agent Training and Evaluation uses one scheduler to coordinate benchmark and agent events across long-running sessions. Consistent evaluation over time helps teams measure whether an agent actually improves, not just whether it passes a static test—Principle 16.

What TLA+ Can and Can’t Check explains that formal verification catches concurrency bugs only when developers specify the properties that matter. Treat proofs as checks against explicit requirements, not a blanket safety stamp for generated code—Principles 02 and 16.