The Inference Report

The infrastructure for autonomous agents is consolidating around two complementary layers: the runtime and the skill. Cloudflare's security-audit-skill and AWS's agent-toolkit-for-aws both treat agent capabilities as modular, independently verifiable components that can be composed into larger workflows. This is a deliberate departure from monolithic agent frameworks. The security-audit-skill enforces machine-readable findings that can be independently verified, which matters because agents making consequential decisions need auditability built in, not bolted on after. NVIDIA's OpenShell takes the other half of the problem: the runtime itself needs to be sandboxed and private by design, not as an afterthought. These repos are gaining traction because they solve a real constraint: developers need agents to do work on production systems without treating the entire agent as a black box or a security liability.

Documentation and template libraries are emerging as a separate category of traction. Nutlope's hallmark and luongnv89's claude-howto are not tools for building agents; they are tools for building better prompts and skills for agents. The distinction matters. These repos succeed because they acknowledge that agent capability lives partly in code and partly in instruction, and that instruction quality compounds across thousands of deployments. Google-gemini's gemini-cli follows a simpler pattern: bring the model into the environment where developers already work. The discovery tier shows emerging specialization: MemWhale and remnic both solve the same problem (persistent, retrievable context for agents) but remnic adds provenance and correction boundaries, suggesting the market is differentiating on which aspects of memory matter most. ZenML and mljar-supervised represent the older pattern, where orchestration and AutoML pipelines are separate concerns. The newer pattern treats agent skills and memory as first-class infrastructure, not as features layered on top of existing ML platforms.

Jack Ridley