KDCube
← AI Industry News

Agent Runtimes Eat the Framework as MCP Cutover Nears

Amazon Bedrock AgentCore's declarative harness hit general availability, letting teams declare a model, tools, and instructions while the runtime handles orchestration, memory, and recovery — as LangGraph 1.0 folds MCP tools into the graph and Netzilo ships cross-platform runtime governance. Meanwhile ...

Highlights

  • Amazon Bedrock AgentCore's harness is now generally available, letting teams "declare what your agent does" — model, tools, instructions — and hand orchestration, memory, and error recovery to the runtime; AWS says agent tasks on AgentCore grew 15x in six months (Amazon)
  • The MCP 2026-07-28 final spec is now 9 days out; the release candidate (locked May 21) does more than go stateless — it adds a Tasks extension and MCP Apps server-rendered UIs via a new extensions framework (MCP Blog)
  • LangGraph 1.0 now treats MCP tools as first-class graph nodes with full streaming, folding the connector layer directly into the orchestration graph (LangChain)
  • Netzilo shipped runtime "Bring Your Own Governance," isolating or terminating a compromised agent in real time across AgentCore, Copilot Studio, CrewAI, LangGraph, and Vertex (Help Net Security)

Key Signals

  1. AWS Summit NYC 2026 AgentCore harness hits GA — the runtime, not the framework, is the product

    AWS made its declarative harness generally available: you specify the model, tools, and instructions, and AgentCore handles orchestration, tool execution, memory, context, and error recovery — plus a managed RAG knowledge base and Guardrails integration. For teams running many agentic apps, this collapses the framework-plus-glue stack into a managed surface, and the 15x task-growth figure (with Nasdaq, Visa, and Experian named) signals real production load, not preview traffic (Amazon).

  2. July 28 final, RC locked May 21 MCP's real story is now the extension surface, not just stateless transport

    Prior coverage focused on killing Mcp-Session-Id so servers scale behind a round-robin load balancer. The RC also promotes Tasks to an extension (servers return task handles; clients drive long-running work) and adds MCP Apps for sandboxed server-rendered HTML. Builders porting for July 28 should scope both the stateless rewrite and whether these extensions change their client contracts (MCP Blog).

  3. July 1 Governance is being rebuilt as a runtime layer that follows the agent

    Netzilo's pitch — "governance cannot depend on which platform exposes which integration point" — targets the fragmentation of per-platform telemetry and policy hooks with Governance-as-Code and real-time kill switches spanning six-plus harnesses. As agents spread across AgentCore, Copilot Studio, and LangGraph in one org, portable runtime enforcement becomes the operator's control plane (Help Net Security).

Why It Matters / What To Watch

  1. The build-vs-buy line is moving up the stack
    • Benchmark AgentCore's declarative harness against your hand-rolled framework loop for a real workload — memory, tool retries, and cost per task, not a demo (Amazon).
    • If you stay on a framework, LangGraph 1.0's first-class MCP nodes reduce the connector glue that used to justify a managed runtime — reassess where your abstraction actually earns its keep (LangChain).
  2. Two clocks are running: MCP July 28 and cross-platform governance
    • Nine days out — confirm your MCP servers drop session affinity, and test the RC's Tasks and MCP Apps extensions against your clients now, not after the cutover (MCP Blog).
    • As agents multiply across vendors, watch whether runtime-governance layers like Netzilo's become table stakes for audit and isolation, or get absorbed into each platform's native controls (Help Net Security).

Quick Links