Fixing Agents: security, traces, deployment, evaluation, orchestration

VulnHunter: Capital One’s agentic AI code security tool — Capital One open-sources VulnHunter, an agentic system that proactively finds exploitable code defects and suggests targeted remediations. Outcome engineers can study its agent-driven scanning + remediation loop as a blueprint for baking continuous, automated security checks into delivery pipelines (Principle 03 & 14).

Your traces already know how to fix your prompt — Pydantic shows production traces can automatically diagnose prompt failures, propose a single evidence-cited edit, and push that edit behind a human review gate. That pattern turns runtime telemetry into a closed-loop prompt management flow you can operationalize for managed prompts and human-in-the-loop validation (Principles 02, 13, 15).

Setting up your spare Mac for Claude Code to control, a step-by-step guide — Yk Dojo documents how to turn a spare Mac into a locked-down Claude Code agent host for remote control while isolating local data. Use this as a practical reference for agent deployment hygiene and host isolation when you need low-cost, auditable execution islands for autonomous agents (Principles 07 & 10).

Show Me Examples: Inferring Visual Concepts from Image Sets — Apple’s VICIS benchmark exposes vision-language models’ failures to infer shared visual concepts from few-shot image sets. That gives outcome engineers a concrete evaluation target and test-suite idea for validating concept-level generalization in perception stacks before agents act on visual inputs (Principle 16).

Feathery raises $30M to rewire financial-services workflows — Feathery secures $30M to scale an AI operating and decisioning platform aimed at rewiring financial workflows with agent orchestration. Watch its approach for how to design decisioning primitives, human approval gates, and orchestration patterns you can reuse when turning agentic prototypes into regulated, outcome-driven services (Principles 04 & 09).