Claude Managed Agents alternatives: 8 options in 2026

Anthropic describes Claude Managed Agents as a suite of composable APIs for building and deploying cloud-hosted agents. It provides the harness: the agent loop runs on Anthropic's servers with prompt caching and compaction built in, and each session gets a fresh Linux container with eight built-in tools for shell commands, files, and web search. Agents connect to remote MCP servers, and vaults keep credentials out of the sandbox: code sees a placeholder, and the real secret is added only to outgoing requests.

It is a developer product. Your application starts sessions, runs custom tools, answers tool calls that pause for permission, and builds the interface people use. Anthropic's quickstart renders a chat surface, and a Chat SDK adapter moves the same handler to Slack or Teams. Scheduled deployments, in public beta, add cron schedules. Pricing is model tokens plus $0.08 per session-hour while a session is running, according to Anthropic's pricing page.

The usual reasons to compare alternatives: you need models other than Claude, you want triggers and a team interface without building them, your company has committed to AWS, Azure, or Google Cloud, or the people who will use the agent are your own team rather than your product's users.

The options fall into three groups. Agent APIs from AI labs, such as Claude Managed Agents, the OpenAI Agents API, and Gemini API Managed Agents, host a harness and a sandbox for developers to build products on. Cloud agent platforms, such as Amazon Bedrock AgentCore, Microsoft Foundry Agent Service, and Google Agent Runtime, give developers the services to deploy and operate agents inside the cloud they already use. Managed platforms for a team's work, such as type.com, Grep, and Duet, add triggers, connections, and an interface, so a team describes the work and the platform runs it.

No group is better in general. An agent API is the right base for an agent inside your product, a cloud platform is the right base for engineers who already run everything on that cloud, and a managed platform is the right choice when the people using the agent are your own team.

Model usage is extra everywhere except where a plan includes it: type.com plans include AI usage metered at the provider's API rates, and Grep and Duet sell credits. Where a column says none are documented, agents start from API calls you make or from infrastructure you set up.

You describe the job in a Space, or write it in TypeScript and deploy it with the type.com CLI. Automations start from schedules (every minute to once a month), Slack messages, GitHub and Linear events, webhooks, emails sent to a Space, or RSS items. Each run makes a thread with the agent's messages and tool calls that the team can read and continue, and run history records the output and any errors. How type.com automations work covers a run from trigger to result.

Connections are added once, through OAuth sign-in, API keys, service accounts, or hosted MCP servers, and assigned to the Spaces that need them. Each run gets a short-lived token, so automation code never holds a provider credential, and scheduled runs act as the person who created the automation. Code can do the fixed steps and hand judgment to the agent in the same automation. Code automations are in early access.

Plans start at $50 a month for 2 team members with $100 of AI usage, metered at the provider's API rates, and team members can connect their own Claude or ChatGPT subscriptions. See type.com pricing and the full type.com vs Claude Managed Agents comparison.

Amazon Bedrock AgentCore became generally available in October 2025. Its managed harness, generally available since June 2026, runs an agent you declare in configuration on a loop built on Strands Agents. Or you write the loop in LangGraph, CrewAI, Google ADK, the OpenAI Agents SDK, or another framework and deploy it to Runtime, where each session gets its own microVM for up to 8 hours. Gateway turns APIs, Lambda functions, and MCP servers into tools, Identity keeps credentials in a token vault, and Policy checks each tool call.

Agents start when your code calls the API, including from AWS Step Functions, and the documentation doesn't describe a built-in scheduler. AWS points business users to Amazon Quick, a separate product. Pricing is consumption-based: Runtime is $0.0895 per vCPU-hour and $0.00945 per GB-hour, billed per second, plus model usage, according to AgentCore pricing. See type.com vs Amazon Bedrock AgentCore for the full comparison.

The Agents API gives your application access to the Codex harness through an OpenAI-managed API. OpenAI manages sessions, orchestration, context compaction, and recovery. The agent runs in an OpenAI-hosted Linux sandbox, in a sandbox you host with a provider such as Modal or Daytona, or with no sandbox. It supports MCP servers, web search, computer use, and vaults that give the sandbox a placeholder instead of the secret.

Your application server submits tasks, handles function tools, and supplies the interface and the triggers; OpenAI's own Slack example runs Slack access and credentials in your app. Model usage is billed at each model's API rates, plus standard rates for tools and OpenAI-hosted containers. Agent Builder, OpenAI's visual builder, shuts down on November 30, 2026, and OpenAI points no-code builders to ChatGPT workspace agents.

Foundry Agent Service is Microsoft's generally available platform for building, deploying, and scaling agents. Prompt agents are defined entirely in configuration. Hosted agents, generally available since July 2026, run your own code from Agent Framework, LangGraph, the OpenAI Agents SDK, Anthropic's agent SDK, or a framework of your own, each in per-session VM-isolated sandboxes with its own Microsoft Entra identity. Models include Azure OpenAI and other Foundry models, and Claude works through Microsoft Agent Framework.

Routines, generally available since September 2026, start agents on a cron schedule as often as every five minutes, on a timer, from a GitHub issue, or from a Teams channel message, and agents can be published to Teams and Microsoft 365 Copilot. Prompt agents carry no extra charge, hosted agents bill container compute from $0.0994 per vCPU-hour, and models are billed separately.

Google renamed Agent Engine to Agent Runtime on April 22, 2026, when Agent Builder became part of Gemini Enterprise Agent Platform. You deploy agent code built with ADK, LangChain, LangGraph, AG2, LlamaIndex, or a custom container, and Google runs and scales it, with per-agent identities, IAM, and an Agent Gateway that governs tool calls. Agents can call Gemini or partner models such as Claude.

Agents start from API calls, and the documentation doesn't describe a built-in cron trigger. Pricing is $0.085 per vCPU-hour and $0.009 per GiB-hour, idle time between turns isn't billed, and the first 50 vCPU-hours and 100 GiB-hours each month are free.

Managed agents in the Gemini API give you a configurable agent harness. One API call provisions a Google-hosted Linux sandbox where the agent reasons, runs code, and manages files, and you define the agent in Markdown files such as AGENTS.md and SKILL.md. Triggers run an agent on a cron schedule, credentials stay outside the agent's environment, and the agent can use remote MCP servers.

Models are Gemini only, and an agent's model is fixed when you create it. Sandbox compute isn't billed during the preview, so you pay Gemini token rates.

In Grep, you describe the job in plain English or a few lines of code. Grep runs each job in a fresh sandbox with research data sources and skills, and chooses Claude, GPT, DeepSeek, or open-source models for each step. Agents run on schedules, as monitors that watch for changes, or in batches, and results arrive in an inbox or by email.

Pricing is credit-based. Pro is $200 a month for 1,500 credits, about 75 medium tasks, and Ultra is $500 a month for 4,500 credits, according to Grep's pricing page.

Duet gives each workspace a dedicated single-tenant VM, keeps models behind one gateway (Claude, GPT, and Kimi models, chosen by quality tier), and handles sign-in and billing. You maintain the agent's instructions, skill files, and memory. Agents run on schedules, wake on events, and listen for webhooks, and people you invite can talk to every agent in the workspace.

Plans are $20, $50, or $200 a month, each with unlimited seats, according to Duet's pricing page.

Per Anthropic's Agent SDK hosting guide, the SDK runs a claude CLI subprocess with its own shell, working directory, and session files, so each agent needs a container. Anthropic suggests 1 GiB of RAM, 5 GiB of disk, and 1 CPU per agent as a starting point, and estimates a minimally provisioned container at roughly $0.05 an hour before tokens.

Everything around the loop is yours: keeping tool credentials out of the agent's environment, adding timeouts, starting runs on a schedule or an event, and building the interface people use. That is the work a managed agent platform does for you.

Most teams can rule out half the options with the first question alone.