10 AI Agent Platforms Businesses Are Evaluating in 2026

Every vendor now has an 'agent'. The word covers a chat assistant that suggests a reply, a framework that lets a developer wire a model to company systems, and a hosted product that handles customer conversations on its own. Putting them in one list is useful only if we are clear about what differs between them.

What does 'autonomous' actually mean for an agent?

An agent, as our guide to AI coding agents puts it, plans its own steps and calls tools instead of answering a single prompt. Autonomy is how far that goes before a person steps in. Three questions place any product on the spectrum:

1. What can it touch? Read-only data, or systems where it can change records, send messages or spend money.
2. What needs approval? Nothing, a list of named actions, or everything.
3. What is recorded? Whether you can reconstruct, afterwards, what it did and why.

Which ten platforms are in this list?

1. LangGraph

The LangGraph overview describes it as a low-level orchestration framework and runtime for 'building, managing, and deploying long-running, stateful agents', usable with or without LangChain. It lists human-in-the-loop (people can review and change an agent's state during a run), persistence and durable execution among its capabilities. Caveat: 'low-level' means your team builds the approval screens and the monitoring; the framework does not supply a finished product.

2. CrewAI

CrewAI calls itself an open-source framework for orchestrating autonomous agents. It pairs Flows, which manage state and control execution, with Crews, which are teams of agents that take on tasks a Flow hands them. Caveat: the introduction page we read does not describe human review or approval features, so a buyer must find and test them in the rest of the documentation. The page also describes itself as 'the leading' framework, a claim we have not assessed and do not repeat as fact.

3. Microsoft Agent Framework

Microsoft's overview presents Agent Framework as a way to build agents and graph-based workflows in .NET, Python and Go. Its 'Harness Agent' lists planning, file access and memory, 'don't-ask-again tool approval' and observability, and the page says its state system supports long-running and human-in-the-loop scenarios. It describes itself as the direct successor to both Semantic Kernel and AutoGen. Caveat: the page says the Go version is in public preview, and tells users that applications built with third-party systems are their own responsibility to review and test.

4. AutoGen

AutoGen's documentation describes 'a framework for building AI agents and applications' in four parts: AgentChat for conversational single and multi-agent apps, Core for event-driven multi-agent systems, Extensions, and Studio, a web interface for prototyping without code. Caveat: Microsoft's Agent Framework page calls itself the direct successor to AutoGen, while the AutoGen page we read states no maintenance status. Check the project's current status before starting a new build on it.

5. OpenAI Agents SDK

The SDK is described as a lightweight Python framework with 'very few abstractions'. Its core ideas are agents (a model given instructions and tools), handoffs or agents-as-tools for delegation, and guardrails that validate inputs and outputs and can stop a run when a check fails. The docs mention built-in human-in-the-loop mechanisms and tracing. Caveat: guardrails check what you tell them to check. They catch the failures you thought of, not the ones you did not.

6. Claude Agent SDK

Anthropic's docs describe the Agent SDK as Claude Code's tools, agent loop and context management as a library for Python and TypeScript. Its capability list includes permissions ('control which tools run automatically, which need approval') and hooks that run your own code at points in the agent's lifecycle. Caveat: the page notes that third-party products may not offer claude.ai login or rate limits unless previously approved and must use API-key authentication. Read the licence and terms section before building a product on it.

7. Google Agent Development Kit (ADK)

ADK is described as an open-source framework to 'build, debug, and deploy reliable AI agents at enterprise scale', available in Python, TypeScript, Go, Java and Kotlin. Its navigation lists graph-workflow human input, function-tool action confirmations and an evaluation section. Caveat: the landing page does not explain those features, and we did not read their pages, so we cannot say how they behave. Treat the list as a prompt for your own check.

8. LlamaIndex agents

LlamaIndex describes an agent as 'a semi-autonomous piece of software powered by an LLM' that chooses from a set of tools at each step and decides after each whether the task is finished. Tools range from plain Python functions to its own query engines. Caveat: the page we read names a human-in-the-loop section but does not describe it. Note the framework's own word: 'semi-autonomous'.

9. Amazon Bedrock Agents (Classic)

AWS describes Bedrock Agents as a service for building agents that break requests into steps, collect missing information from the user, call company APIs and query knowledge bases, with traces to examine the agent's reasoning at each step. Caveat: the AWS page states that Bedrock Agents (now 'Bedrock Agents Classic') is no longer open to new customers, that existing customers can continue as normal, and points to Amazon Bedrock AgentCore for similar capabilities. We did not read the AgentCore documentation, so we do not describe it.

10. Salesforce Agentforce

Salesforce presents Agentforce as an AI agent platform for customer and employee service across web, phone and apps. Its page says agents work 'within set guardrails', draw on business data such as CRM data, and can 'escalate the matter to human agents' when an issue is out of scope. Caveat: this is vendor marketing. It describes intended behaviour, not measured results, and we state none of its performance claims. Ask for evidence under your own conditions.

What governance questions should you ask before trusting any of them?

The platform matters less than the controls you set around it. Use the same evidence discipline as in our procurement evidence checklist and vendor claims versus verified findings guides.

What are the limitations of this list?

We read one or two documentation pages per product on a single day. Docs change, features move between versions, and a page can omit a feature that exists. We did not install, run or test any of these tools, and nothing here is a performance comparison. The list is not exhaustive and its order is not a ranking. Confirm every detail on the vendor's current documentation before you decide.