KDCube
← AI Industry News

Policy-as-Code Slides Into the Agent SDK

Arcjet launched agent runtime security on Sept 17 — an Open Policy Agent layer that checks each action before it runs and re-verifies the result after, embedded across the Claude Agent SDK, OpenAI Agents SDK, LangChain, Mastra, and Microsoft Agent Framework. Alongside it, inference engines vLLM and SGLang tuned for tool-heavy ...

AI Industry Dossier

Highlights

  • Arcjet launched agent runtime security on Sept 17 — an Open Policy Agent-based layer that checks every agent action before it runs and re-verifies the result after, embedded directly in the Claude Agent SDK, OpenAI Agents SDK, LangChain, Mastra, and Microsoft Agent Framework. (SiliconANGLE)
  • The inference tier keeps racing on agentic serving: vLLM's v0.28.0 line pushed DeepSeek-V4.1 and disaggregated KV transfer, while SGLang's 0.5.x cadence leaned into DFlash and decode context parallelism. (vLLM, RunAnywhere)
  • UiPath reframed its Integration Service as an "agent-ready connectivity layer" — credentials, actions, and event sources that coding agents build against but UiPath governs. (UiPath)
  • Data-control gateways went per-call: Dipp AI shipped runtime enforcement of training-exclusion and residency boundaries on every agent request. (AI Agent Store)

Key Signals

  1. Runtime security converges on policy-as-code - Sept 17, 2026

    Arcjet's launch moves the agent-security conversation from perimeter tooling toward rules-as-code that live beside the model call. Policies run on Open Policy Agent and can be changed without redeploying the app; prompt-injection detection and PII redaction ship as built-ins, and correlated traces log inputs, actions, and policy evaluations per run. Python and Go SDKs hit 1.0 alongside the release, with 500+ production apps cited. (SiliconANGLE)

  2. The "before-and-after" check becomes a pattern - Sept 17, 2026

    Arcjet's dual-stage model — evaluate the intended action, then verify the actual result — is the operationally interesting bit. It targets the gap the Sept 14 sandbox-escape stories exposed: an agent can be authorized to act yet still produce a harmful outcome. Expect framework-agnostic, SDK-embedded guardrails to become table stakes, not a bolt-on. (SiliconANGLE)

  3. Serving engines optimize for agents, not just chat - Sept 2026

    vLLM and SGLang are tuning for long, tool-heavy agent loops: disaggregated KV transfer, decode context parallelism, and speculative decoding aimed at throughput under concurrency rather than single-turn latency. For operators, the model runtime and the agent runtime are co-evolving. (vLLM, RunAnywhere)

Why It Matters / What To Watch

  1. Guardrails are moving into the SDK layer — audit where yours live.
    • If your policy engine is a separate proxy, compare against inline, OPA-driven checks that travel with the agent call and log per-run traces. (SiliconANGLE)
    • Watch for the post-execution verification step; authorizing an action is no longer the same as approving its result. (SiliconANGLE)
  2. Connectors and data boundaries are becoming governed agent surfaces.
    • Treat integration layers as agent identity/credential surfaces — UiPath now positions Integration Service as the governed layer agents build through, not around. (UiPath)
    • Evaluate per-call data-control gateways (e.g., Dipp AI) if training-exclusion or residency is a compliance requirement. (AI Agent Store)
  3. Match inference upgrades to agentic workloads.
    • Re-benchmark vLLM/SGLang for concurrent, tool-calling loops rather than chat latency before your next serving decision. (vLLM, RunAnywhere)

Quick Links