Published October 30, 2026
AI agent orchestration software lets a large language model plan a task, call tools and hand work to other agents, and keeps track of each step. Frameworks supply the code that defines agents and how they work together. Runtimes run those agents in production and keep their state, identity and logs.
Where the market stands
Anthropic released the Model Context Protocol (MCP) as an open standard on November 25, 2024. MCP connects AI applications to tools and data sources. The current protocol version is 2026-07-28.
Google launched the Agent2Agent protocol (A2A) in April 2025 for communication between agents. The Linux Foundation launched the A2A project on June 23, 2025. By April 9, 2026, A2A had released version 1.0, its first stable specification, and the Linux Foundation counted more than 150 supporting organizations.
On December 9, 2025, Anthropic donated MCP to the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI. A2A joined the AAIF as a hosted project on August 17, 2026.
OWASP published the Top 10 for Agentic Applications 2026, a framework of security risks for AI agents, on December 9, 2025.
Three kinds of product
The products fall into three kinds. One deployment can combine more than one kind.
- Open-source frameworks. Code libraries for defining agents, tools and the flow between them. They include LangGraph, CrewAI, Google's Agent Development Kit (ADK), Microsoft Agent Framework, the OpenAI Agents SDK and LlamaIndex, all open source. ADK is available in Python, TypeScript, Go, Java and Kotlin. The OpenAI Agents SDK also supports non-OpenAI model providers.
- Managed runtimes. Cloud services that host and run agents and keep their state. They include Amazon Bedrock AgentCore, Agent Runtime in Google's Gemini Enterprise Agent Platform, Microsoft Foundry Agent Service and the OpenAI Agents API. AWS says AgentCore works with any open-source framework and with models inside and outside Amazon Bedrock.
- Durable execution engines. Software that saves a workflow's state so it resumes after a failure instead of starting over. Temporal is an MIT-licensed durable execution platform. Its integration runs OpenAI Agents SDK agents as Temporal Workflows. Model calls retry durably, and agents survive Worker restarts without losing state.
Support for the two protocols varies by product. AgentCore Runtime supports A2A, and AgentCore Gateway connects to existing MCP servers. Microsoft Agent Framework 1.0 supports MCP. On April 28, 2026, Microsoft added A2A version 1 support to the framework's .NET packages, which it says are still in preview.
What changed at the vendors
- LangChain. LangChain 1.0 and LangGraph 1.0 shipped on October 22, 2025, with a commitment to no breaking changes until 2.0. LangGraph 1.0 includes durable state, built-in persistence and human-in-the-loop patterns. On October 20, 2025, LangChain announced it had raised $125 million at a $1.25 billion valuation, in a round led by IVP.
- AWS. Amazon Bedrock AgentCore became generally available on October 13, 2025, with VPC and AWS PrivateLink support across its services.
- Microsoft. Agent Framework reached version 1.0 for .NET and Python on April 3, 2026. Microsoft says it unifies the foundations of Semantic Kernel with the orchestrations of AutoGen, and the framework is open source. The AutoGen repository is archived. Microsoft says it no longer maintains AutoGen and directs new users to Agent Framework. Microsoft publishes migration guides from both predecessors.
- Google. Google announced Gemini Enterprise Agent Platform on April 22, 2026. It brings together Vertex AI's model selection, model building and tuning services with new agent features, including Agent Registry, Agent Identity and Agent Gateway. ADK gained a graph-based framework for defining how agents work together.
- OpenAI. The Assistants API shut down on August 26, 2026, replaced by the Responses API and Conversations API. On June 3, 2026, OpenAI announced the deprecation of Agent Builder, its visual workflow canvas, and set its shutdown for November 30, 2026. OpenAI points Agent Builder users to the Agents SDK or ChatGPT Workspace Agents. ChatKit remains available. The Agents API, which runs OpenAI's Codex harness as a managed service, entered public beta on September 10, 2026.
What each costs where prices are public
CrewAI, LlamaIndex and Temporal release their open-source code under the MIT License. The paid tiers are hosted services and runtimes. Figures are list prices as published on October 17, 2026.
- LangSmith (LangChain). The Developer plan is $0 per seat per month with up to 5,000 base traces a month. Plus is $39 per seat per month with up to 10,000 base traces. Traces beyond that are pay as you go. Enterprise is custom. Deployments are metered in LangChain Standard Units at $1.00 each, with runtime compute at 0.0675 units per vCPU-hour.
- CrewAI. The Basic plan is free and includes 50 workflow executions a month. Enterprise pricing is custom.
- LlamaCloud (LlamaIndex). Usage is billed in credits at $1.25 per 1,000. Free includes 10,000 credits, Starter costs $50 a month with 40,000 and Pro costs $500 a month with 400,000. Basic parsing costs from 1 credit per page.
- Temporal Cloud. Pay-as-you-go starts at $50 per million actions and falls to $25 per million with volume. Business starts at $500 a month or 10% of usage, whichever is higher. Active storage is $0.042 per GB-hour.
- Amazon Bedrock AgentCore. Runtime on microVMs v2 with consumption billing costs $0.1276 per vCPU-hour and $0.0169 per GB-hour, billed per second with a 1-second minimum. CPU is not billed while an agent waits on a model, tool or database. AWS also lists v1 rates of $0.0895 per vCPU-hour and $0.00945 per GB-hour, and a lower committed-baseline rate for v2 that it says is launching by October 2026. Gateway costs $0.005 per 1,000 invocations.
- Agent Runtime (Gemini Enterprise Agent Platform). Google bills Agent Compute at $0.085 per vCPU-hour and Agent Memory at $0.009 per GiB-hour, in USD. The first 50 vCPU-hours and 100 GiB-hours per account each month are free. Idle time between turns is not billed.
- Microsoft Foundry Agent Service. No additional charge for agents built from prompts and workflows. Model tokens, tools and connectors are billed separately. Hosted agents that run custom code are billed on container compute per hour.
- OpenAI Agents API. Model usage is billed at the model's API rates, tools at their standard rates and hosted sandboxes at container rates.
Price model usage separately from the runtime. Estimate model calls per run, including retries and calls between agents, before comparing platforms.
What a deployment requires
The MCP specification says hosts must obtain explicit user consent before invoking any tool. It says descriptions of tool behavior should be treated as untrusted unless they come from a trusted server.
Plan for tracing, cost limits and human approval from the first agent. Record every model call and tool call with its cost. Set a token budget per agent and an alert for runaway loops. Pause for human approval before any action that is costly to reverse.
Test the failure path in the pilot, not only the working path. Stop the process mid-run, throttle the model API and pass a prompt-injection payload through a tool result. Check whether the platform shows where the run stopped, resumes from that point and blocks the injection.
Start with one narrow task and the fewest agents that complete it. Add agents and autonomy as tracing and test results show the first one is reliable.
Questions to settle first
Settle these before choosing a platform.
- Who will run the agent runtime: your team, a cloud provider or a model vendor?
- Which steps must survive a crash or timeout without being repeated, and what will save their state?
- Do the tools and systems your agents need already offer MCP servers, and will agents from other vendors need to reach yours over A2A?
- If you built on OpenAI's Assistants API or Agent Builder, AutoGen or Semantic Kernel, what is the migration plan and date?
- What does the pilot have to prove about tracing, cost per run and failure recovery before a second agent is added?