DN Managed Agent Runtime Index 2027

The Best Infrastructure for Running AI Agents

OpenAI Agents API vs Amazon Bedrock Managed Agents vs Google Vertex AI Agent Engine vs Microsoft Foundry and Agent Framework vs self-hosted stacks.

What Matters

Google Vertex AI Agent Engine leads this edition by one point because it combines a mature managed runtime with memory, sessions, code execution, observability, identity controls and broad framework support. OpenAI Agents API is the fastest route to the Codex harness and ranks first for developer velocity, but its public-beta status and narrower enterprise control surface keep it just behind.

Microsoft is the strongest choice for enterprises already committed to Azure and Microsoft 365. Amazon Bedrock Managed Agents offers the most direct route to OpenAI agents that run inside AWS, although its preview status is a material production constraint. Self-hosted frameworks remain the control and portability winner, but teams pay for that freedom through engineering, security and reliability work.

The 2027 ranking

Important: These are evidence-based readiness scores, not claims of measured latency or task-completion superiority. A two-point difference is not statistically meaningful. Readers should use the adjustable tool below to model their priorities.

DN Evidence Block

Question: Which runtime is most ready to operate persistent, tool-using AI agents in production?

Evidence window: Public product documentation and announcements available through 8 October 2026.

Included: Managed execution, state and memory, tools and protocols, security, observability, deployment, model portability, operational control and price transparency.

Not yet included: DN-operated latency, recovery, adversarial-security or cost-per-successful-task tests. Vendor testimonials are not treated as independent benchmark results.

What is a managed agent runtime?

A model API generates outputs. An agent runtime keeps a goal-directed system operating across many model calls, tool invocations and interruptions. It may preserve session state, compact context, retry failed actions, isolate code, coordinate specialist agents, store memories, trace decisions and enforce identity or permission policies.

This distinction matters because an apparently more intelligent agent may actually be benefiting from a better harness. Context management, tool routing, recovery and sandbox design can change real-world task completion without any change to the underlying model.

The 2027 infrastructure contest is therefore not simply OpenAI versus Gemini or one foundation model versus another. It is a contest among complete execution systems.

DN Runtime Fit Calculator

1. Google Vertex AI Agent Engine: best overall

Google's managed agent layer, increasingly presented within the Gemini Enterprise Agent Platform, offers a serverless runtime plus Sessions, Memory Bank, code execution, tracing, logging, monitoring and identity controls. Its documentation supports Google's Agent Development Kit as well as LangChain, LangGraph, LlamaIndex, AG2 and A2A agents.

That breadth is strategically important. A company can use Google infrastructure without committing every orchestration decision to a Google-only framework. Agent identity, service accounts, IAM conditions, private connectivity and Agent Gateway also give security teams recognizable control points.

Why it leads

• Broad framework and A2A support.
• Managed sessions and long-term memory.
• Integrated tracing, logging, monitoring and evaluation.
• Secure managed code execution and computer-use direction.
• Clearer component pricing than several rivals.

Where it can disappoint

Capability breadth creates configuration overhead. Costs can also fragment across runtime compute, memory, stored session events, Memory Bank retrieval, models and tools. The platform is most compelling for teams already equipped to manage Google Cloud IAM, billing and network architecture.

2. OpenAI Agents API: best for developer velocity

The Agents API exposes the Codex harness as an OpenAI-managed service. OpenAI manages sessions, orchestration, context compaction and recovery. Agents can execute code, edit files, use custom functions, search the web, connect to MCP servers and operate in hosted or partner execution environments.

Its most important advantage is integration. Long-running sessions, tool search, programmatic tool calling, subagents and computer use are parts of one opinionated system rather than an assembly project.

Why it nearly wins

• Fast path from prototype to durable cloud agent.
• Open-source Codex harness provides visibility into core orchestration logic.
• MCP, custom tools, web search, computer use and subagents.
• Choice of OpenAI-hosted, customer-managed or partner sandbox environments.
• No separate Agents API fee during the reviewed public-beta offer; tokens and paid tools remain billable.

The trade-off

The service was still in public beta at the evidence cut-off. OpenAI manages important orchestration behavior, which accelerates delivery but reduces direct operational control. Organizations should verify data residency, support, quotas, tool isolation and migration requirements before treating it as a default for regulated workloads.

3. Microsoft Agent Framework and Foundry: best Microsoft ecosystem fit

Microsoft's offering has two related layers. Microsoft Agent Framework is the open-source successor to AutoGen and Semantic Kernel, combining multi-agent abstractions with state management, type safety, middleware, telemetry and deterministic workflows. Foundry Agent Service adds hosted agents, managed compute, tracing, evaluation, identity and publishing into Microsoft channels.

The framework supports sequential, concurrent, handoff, group-chat and manager-directed orchestration patterns. Graph-based workflows can checkpoint state, survive interruption and incorporate human approval.

Why enterprises will choose it

• Natural fit with Azure identity, networking and governance.
• Open-source framework with Python, .NET and Go support.
• Explicit workflows for deterministic and human-in-the-loop processes.
• Hosted runtime available across a wide regional footprint, including South Africa North.
• Distribution into Teams, Microsoft 365 and custom applications.

Why the score is not higher

Microsoft offers several overlapping paths: the framework, Foundry Agent Service, model providers, hosted agents and Copilot-oriented distribution. That is powerful, but it increases architectural decision load. Pricing also depends on the selected models, hosted compute and tools rather than resolving to one simple agent price.

4. Self-hosted frameworks: best for control

Self-hosting is a category, not one product. It may combine LangGraph, CrewAI, Microsoft Agent Framework or another open framework with Kubernetes, Cloud Run, containers, private inference, external model APIs, custom memory and an observability stack.

The strongest self-hosted architecture can exceed every managed service for data control, customization and portability. The median implementation may be less reliable because its owners must build checkpointing, retry policies, secrets management, isolation, scaling, telemetry and incident response themselves.

Advantages

• Maximum control over infrastructure, models, data and logs.
• Ability to run in private clouds, sovereign environments or on-premises.
• Reduced dependence on a single runtime vendor.
• Freedom to tune orchestration and economics for a narrow workload.

Hidden costs

• Platform engineering and on-call staffing.
• Sandbox and secret-isolation design.
• Version compatibility and framework upgrades.
• Evaluation, tracing and governance integration.
• Capacity planning and recovery testing.

5. Amazon Bedrock Managed Agents: best AWS-native route to OpenAI

Amazon Bedrock Managed Agents is a jointly developed, AWS-native version of OpenAI's Agents API. It runs stateful agents using OpenAI models on Bedrock while integrating with AWS identities, permissions and governance. Developers provide an execution environment, while the service manages the session and agent-model interaction.

This is strategically significant because enterprises can use the Codex harness and OpenAI models within their AWS operating model and apply usage toward existing cloud commitments where applicable.

Why it ranks fifth today

The ranking reflects readiness, not potential. At the evidence cut-off the product was in preview and documentation listed regional preview endpoints in US East and US West locations. Preview status, boundaries and regional availability must be treated as material constraints for production procurement.

Dimension-by-dimension scorecard

How to choose in five questions

1. Where must data and execution live? Residency or existing cloud controls can eliminate options before feature comparison begins.
2. Who owns reliability? If there is no platform team to operate state, queues, sandboxes and recovery, managed infrastructure deserves a large weight.
3. How important is model portability? Do not score a runtime as portable merely because it speaks to external tools. Test whether agent definitions, sessions and policies can move.
4. What is the unit of cost? Compare cost per successful business outcome, not token prices alone.
5. What evidence survives an incident? Production agents need traces, tool receipts, identities, policy decisions and human override records.

The DN Alpha Thesis

The agent-runtime market is likely to split into three economic layers. Foundation-model companies will sell integrated intelligence and harnesses. Cloud platforms will sell governed execution near enterprise data. Open frameworks will become a portability and bargaining layer across both.

The durable winner may therefore not have the highest raw model score. It may be the runtime that makes agents easiest to audit, recover, authorize and move. As agent work becomes longer and more economically consequential, reliability-adjusted cost will matter more than price per million tokens.

This creates a measurement opportunity for Decentralised News: track cost per verified successful task, not model price. A cheap run that fails, requires human cleanup or triggers the wrong tool is economically expensive.

Methodology

DN reviewed official documentation, pricing information, release material and product-status disclosures. Each runtime category received a 0 to 10 score in nine dimensions. Scores were multiplied by the stated default weights and converted to 100 points. Capabilities documented as preview received maturity or availability penalties. Unverified marketing claims and customer testimonials were not counted as independent performance evidence.

The self-hosted score represents a competent production deployment using a modern open framework. It is not a score for every self-hosted implementation. Actual outcomes vary more widely in this category than in the managed categories.

Planned live benchmark: identical multi-step research, coding, document-processing and tool-use workloads; repeated trials; success adjudication; cold and warm latency; recovery after interruption; tool-call accuracy; long-session drift; context-compaction loss; operator minutes; and total cost per successful task.

Primary sources

• OpenAI: Introducing the Agents API
• OpenAI Developers: Agents API overview
• AWS: Bedrock Managed Agents preview announcement
• AWS documentation: Bedrock Managed Agents powered by OpenAI
• Google Cloud: Gemini Enterprise Agent Platform documentation
• Google Cloud: Scaling agents in production
• Microsoft: Agent Framework overview
• Microsoft Foundry: Hosted agents
• CrewAI: Agent Management Platform documentation
• LangChain: LangSmith deployment and data-plane documentation

Frequently asked questions

Disclosure

This article is independent research and general information, not investment, legal, security or procurement advice. Product capabilities, status, availability and prices can change. Decentralised News may earn revenue from clearly disclosed commercial relationships, but no payment purchases inclusion, methodology changes or ranking position. No unverified affiliate links are used in this edition.