KDCube
← AI Industry News

Agent Runtimes Scale Up: AgentCore V2, Microsoft & Akamai Sandboxes

Amazon Bedrock AgentCore Runtime V2 lifts its ceiling to 250 concurrent connections per session and adds an OverheadLatency metric, while Microsoft previews a GitHub-built Foundry Copilot harness and Akamai ships Inference plus a WebAssembly Agent Sandbox ...

AI Industry Dossier · Agent Runtimes

Highlights

  • Amazon Bedrock AgentCore Runtime V2 agents can now hold up to 250 concurrent connections per session on the default 2-vCPU allocation — a jump from the old 16 — and the platform now emits an OverheadLatency metric so operators can separate platform cold-start time from their own agent's processing (AWS Docs)
  • Microsoft opened a preview of the Foundry Agent Service Copilot harness, a GitHub-built runtime that manages the agent loop, sandbox, five built-in tools (including shell), and MCP tool approvals — with metered billing starting Dec 1, 2026 (Futurum)
  • Akamai launched its Inference model-serving service and Agent Sandbox in beta, isolating agent-generated JavaScript and Python inside a WebAssembly-based restricted runtime on Akamai Functions (Akamai)
  • The through-line this morning: the agent runtime — concurrency, latency metrics, code isolation, tool-approval policy — is where the platform fight has moved, not the model (AWS Docs)

Key Signals

  1. AgentCore V2 raises its connection ceiling and exposes platform latencyNovember 2026, AWS release notes

    Connection-heavy agents using WebSocket streams, SSE, or parallel tool calls can now hold roughly 15x more concurrent connections per session, applied automatically to all V2 runtimes with no action required. The new OverheadLatency CloudWatch metric (in AWS/Bedrock-AgentCore) measures from request receipt to your agent's first response byte, excluding your agent's own compute — the kind of instrumentation operators need to tune cold starts at scale (AWS Docs).

  2. Microsoft's Copilot harness keeps tool approval under Microsoft's controlOct 28, 2026

    The optional Foundry runtime bundles reusable skills, authenticated Foundry and MCP tools, file and code execution, and agent definitions stored in source control — removing the need to maintain your own agent loop. Notable caveat: during preview Microsoft publishes no export pathway to framework code (unlike AWS's AgentCore), and compute meters at $0.0994 per vCPU-hour and $0.0118 per GiB-hour from Dec 1 (Futurum).

  3. Akamai bets on WebAssembly isolation for untrusted agent codeOct 27, 2026

    Agent Sandbox confines generated code with capped compute, memory, and external connections, pitched as a safe place to run model output at the edge; the paired Inference service is a managed model-serving beta (Akamai).

Why It Matters / What To Watch

  1. Runtime limits are now a real design input, not an afterthought.
    • If you run streaming or parallel-tool agents on Bedrock, revisit your per-session concurrency assumptions — the V2 ceiling just moved to 250, which may let you consolidate sessions (AWS Docs).
    • Wire the new OverheadLatency metric into dashboards before blaming your own code for slow responses (AWS Docs).
  2. "Managed harness" convenience trades against portability.
    • Weigh Microsoft's no-export-path preview against AgentCore's framework-agnostic runtime before committing agent definitions to one vendor's loop (Futurum).
    • Price the harness honestly: metered vCPU/GiB billing starts Dec 1, so pilot usage won't be free for long (Futurum).
  3. Code isolation is becoming table stakes for agent execution.
    • If you let agents generate and run code, evaluate sandboxing approaches like Akamai's WebAssembly confinement against your own container or VM isolation (Akamai).

Quick Links