KDCube
← AI Industry News

Managed Agent Runtimes Go Mainstream: OpenAI, Salesforce, AWS

OpenAI's Agents API hit public beta, putting the Codex harness — durable sessions, context compaction, subagents, and pluggable sandboxes — behind one call, while Salesforce shipped a long-horizon runtime and open-source Agent Script at Dreamforce. Meanwhile AWS made AgentCore Evaluations framework-agn...

Highlights

  • OpenAI put the Codex harness behind a single API call: the Agents API entered public beta with server-side sessions, automatic context compaction, subagents, MCP/tools, and a choice of OpenAI-hosted, self-hosted, or nine partner sandboxes (MarkTechPost)
  • At Dreamforce, Salesforce shipped a long-horizon runtime for Agentforce — agents that pursue goals over days and weeks via memory, durable execution, and dynamic steering — plus Agent Script, an open-source language for agent behavior (Salesforce)
  • Amazon Bedrock AgentCore Evaluations went framework-agnostic in September, now scoring TypeScript agents built on Strands, LangGraph, OpenAI Agents, and the Vercel AI SDK (AWS)
  • AgentCore also added a Consent Portal to Identity, so end users explicitly approve what an agent may access on their behalf before it proceeds (AWS)

Key Signals

  1. OpenAI's Agents API turns the Codex harness into infrastructure public beta, Sep 10 The managed runtime handles what teams used to hand-roll: durable sessions, context compaction run independently for the root agent and each subagent, tool discovery, and configurable sandbox placement across partners like Cloudflare, Modal, E2B, and Vercel — defaulting to gpt-6-astra. The catch for regulated shops: it's US-only data residency with no Zero Data Retention support today. (MarkTechPost)
  2. Salesforce reframes agents from chats to campaigns Dreamforce 2026 The long-horizon runtime lets an agent like Hunter turn "rescue my at-risk deals this quarter" into a plan with guardrails on when it can act versus when a human must approve. Multi-Agent Orchestration is now GA, and most job-ready agents (Casey, Paige, Carter, Marshall, Piper, Fin) ship GA now, with Hunter in pilot ahead of a November GA. (Salesforce)
  3. AWS makes agent evaluation portable AgentCore release notes, Sep 2026 Evaluations now covers TypeScript framework variants and, via OpenTelemetry/OpenInference conventions, generic instrumented agents — reducing lock-in for teams that want one scoring surface across LangGraph, OpenAI Agents, LlamaIndex, Google ADK, and the Claude Agent SDK. (AWS)

Why It Matters / What To Watch

  1. The runtime is becoming the product — and the lock-in.
    • Weigh OpenAI's managed sessions against portability: the sandbox is pluggable, but data residency and ZDR gaps may push regulated workloads to self-hosting the codex exec-server path. (MarkTechPost)
    • Salesforce's Agent Script being open source is the tell — vendors know builders fear black-box runtimes and are competing on how much control they hand back. (Salesforce)
  2. Governance is shifting from policy docs to runtime primitives.
    • AgentCore's Consent Portal makes user consent a gated API step, not an afterthought — a concrete pattern to copy for any agent touching third-party resources. (AWS)
    • Watch reliability plumbing too: this week Moveworks shipped a model upgrade that returns explicit tool-call failure states instead of silent success, the kind of honest-failure signal long-horizon agents need to retry safely. (AI Agent Store)

Quick Links