---
title: "How Guild compares to other AI agent platforms"
description: "Guild compared with Claude Managed Agents, AWS, LangChain, Runlayer, Arcade, Vercel eve, feature by feature, with every cell sourced and dated. Last verified 2026-10-30."
canonical: "https://www.guild.ai/compare"
---

# How Guild compares to other AI agent platforms

Guild compared with Claude Managed Agents, AWS, LangChain, Runlayer, Arcade, Vercel eve, feature by feature, with every cell sourced and dated. Last verified 2026-10-30.

## Comparison pages

- [Guild vs. Claude Managed Agents](https://www.guild.ai/compare/claude-managed-agents) ([markdown](https://www.guild.ai/compare/claude-managed-agents.md)): Guild vs Claude Managed Agents: sandboxes, tools, credentials, approvals, schedules, models, and pricing compared, plus how to move an agent to Guild.
- [Guild vs. AWS](https://www.guild.ai/compare/aws) ([markdown](https://www.guild.ai/compare/aws.md)): Guild vs Amazon Bedrock AgentCore: runtime, tools, identity, policy, and evaluations compared with Guild's agents, permissions, triggers, Agent Hub, and CLI.
- [Guild vs. LangChain](https://www.guild.ai/compare/langchain) ([markdown](https://www.guild.ai/compare/langchain.md)): Guild vs LangChain: LangGraph, LangSmith deployment, and Fleet compared with Guild's agents, permissions, triggers, Agent Hub, and CLI, plus how to move.
- [Guild vs. Runlayer](https://www.guild.ai/compare/runlayer) ([markdown](https://www.guild.ai/compare/runlayer.md)): Guild vs Runlayer: Runlayer Agents, its MCP gateway, and its rules for AI tools compared with Guild's agents, permissions, triggers, Agent Hub, and CLI.
- [Guild vs. Arcade](https://www.guild.ai/compare/arcade) ([markdown](https://www.guild.ai/compare/arcade.md)): Guild vs Arcade: per-user tools, MCP gateways, and tool evals compared with Guild's agents, triggers, permissions, and Agent Hub, plus how to move Arcade-backed agents.
- [Guild vs. Vercel eve](https://www.guild.ai/compare/vercel-eve) ([markdown](https://www.guild.ai/compare/vercel-eve.md)): Guild vs Vercel eve: building, integrations, credentials, triggers, and approvals compared, plus how to move an eve agent to Guild.

## Where Guild fits

Frameworks and agent platforms each cover part of the job. Guild runs every agent, sets what each one can reach, and shows what it did, on any model.

- **Frameworks**: LangChain, Vercel eve
- **Agent harnesses**: Claude Managed Agents
- **Cloud services**: AWS (Amazon Bedrock AgentCore)

## Feature grid: Guild, Claude Managed Agents, AWS, LangChain

Each cell: status, the note, and its source. The full grid for every platform is at https://www.guild.ai/compare/llms.txt.

### Authoring

| Feature | Claude Managed Agents | AWS | LangChain | Guild |
| --- | --- | --- | --- | --- |
| Code-first TypeScript agents | **Partial.** The TypeScript SDK (client.beta.agents, client.beta.sessions) creates agents as versioned config: model, system prompt, tools, MCP servers, skills. Anthropic runs the loop; your code only runs custom tools. ([source](https://platform.claude.com/docs/en/managed-agents/agent-setup.md), checked 2026-10-30) | **Yes.** Strands Agents TypeScript SDK (@strands-agents/sdk), deployable to AgentCore Runtime. ([source](https://strandsagents.com/docs/user-guide/sdk/quickstart/typescript/index.md), checked 2026-10-30) | **Yes.** LangChain JS createAgent, or LangGraph StateGraph for explicit graphs. ([source](https://docs.langchain.com/oss/javascript/langchain/agents), checked 2026-10-23) | **Yes.** @guildai/agents-sdk: agent(), llmAgent(), and the task API (task.llm, task.tools, task.ui, task.gather). ([source](https://docs.guild.ai/packages/agents-sdk.md), checked 2026-10-23) |
| Python / LangGraph agents | **No.** No LangGraph hosting. Agents are Anthropic's own harness configured through the API; Python appears only as an SDK for calling it. ([source](https://platform.claude.com/docs/en/managed-agents/overview.md), checked 2026-10-30) | **Yes.** AgentCore Runtime hosts LangGraph agents (and other Python frameworks) wrapped in the BedrockAgentCoreApp entrypoint. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html), checked 2026-10-30) | **Yes.** Native: create_agent (Python) and LangGraph, deployed with LangSmith Deployment. ([source](https://docs.langchain.com/langsmith/deployment), checked 2026-10-23) | **Yes.** LangGraph agent type: langgraph.json + graph.py using guild_model() and guild_tools(). LangChain v1, Python 3.13, no extra pip packages. ([source](https://docs.guild.ai/guide/agent-types.md), checked 2026-10-23) |
| Instruction-only (no-code) agents | **Yes.** The Console agent builder sets model, system prompt, MCP servers, tools, and skills, with an inline test session runner. ([source](https://platform.claude.com/docs/en/managed-agents/onboarding.md), checked 2026-10-30) | **Yes.** AgentCore Harness (GA July 2026): CreateHarness takes model, system prompt, tools, and skills as configuration; no orchestration code or container. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/harness.html), checked 2026-10-30) | **Yes.** LangSmith Fleet (formerly Agent Builder): describe the agent in natural language or start from a template. ([source](https://docs.langchain.com/langsmith/fleet), checked 2026-10-23) | **Yes.** Native agents: PROMPT.md + guild.yaml, no container, ready immediately. ([source](https://docs.guild.ai/guide/agent-types.md), checked 2026-10-23) |
| Several agent formats on one platform | **No.** One agent definition (model, system, tools, MCP servers, skills) run by Anthropic's harness. ([source](https://platform.claude.com/docs/en/managed-agents/overview.md), checked 2026-10-30) | **Yes.** AgentCore Runtime runs Strands, LangGraph, CrewAI, LlamaIndex, Google ADK, OpenAI Agents SDK, or custom code, next to config-only Harness agents. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html), checked 2026-10-30) | **Yes.** Agent Server also hosts Google ADK, Claude Agent SDK, CrewAI, and Strands agents. ([source](https://docs.langchain.com/langsmith/deployment), checked 2026-10-23) | **Yes.** Native, TypeScript, Goose, OpenClaw, and LangGraph agent types, plus ADK (CLI 0.28). ([source](https://docs.guild.ai/guide/agent-types.md), checked 2026-10-23) |
| Typed agent input/output | **Partial.** Custom tools declare a JSON Schema input_schema, and outcomes grade a deliverable against a rubric. No schema for the agent's own input or output is documented. ([source](https://platform.claude.com/docs/en/managed-agents/tools.md), checked 2026-10-30) | **Partial.** Strands structured output returns typed results in code; AgentCore Runtime invocations take a free-form payload with no declared agent schema. ([source](https://strandsagents.com/docs/user-guide/sdk/agents/structured-output/index.md), checked 2026-10-30) | **Yes.** responseFormat for structured output and contextSchema for typed context. ([source](https://docs.langchain.com/oss/javascript/langchain/agents), checked 2026-10-23) | **Yes.** Zod inputSchema and outputSchema on agent(); root input must be z.object. ([source](https://docs.guild.ai/packages/agents-sdk.md), checked 2026-10-23) |
| Repo / workspace agents (shell, git, files) | **Yes.** Each session gets its own Linux container with bash and read, write, edit, glob, and grep tools; GitHub repositories can be mounted, with the GitHub MCP for pull requests. ([source](https://platform.claude.com/docs/en/managed-agents/github.md), checked 2026-10-30) | **Yes.** AgentCore Harness gives each session its own microVM with a filesystem, shell, and file tools; bring your own container image. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/harness.html), checked 2026-10-30) | **Partial.** Deep Agents harness (Fleet is built on it); sandboxing depends on your deployment. ([source](https://docs.langchain.com/langsmith/fleet), checked 2026-10-23) | **Yes.** OpenClaw agents (AGENTS.md + skills/) run in the guildai~lobsterpot container with shell, git, and file editing. ([source](https://docs.guild.ai/guide/agent-types.md), checked 2026-10-23) |
| Sub-agents and agent-to-agent calls | **Yes.** Multiagent orchestration: a roster of predefined or inline agents the session delegates to on separate threads, plus dynamic workflows for large fan-out work and an optional advisor model. ([source](https://platform.claude.com/docs/en/managed-agents/multiagent-orchestration.md), checked 2026-10-30) | **Yes.** Strands agents-as-tools, swarm, graph, and A2A; AgentCore Runtime hosts A2A servers. Harness itself does not do graph or multi-agent patterns. ([source](https://strandsagents.com/docs/user-guide/sdk/multi-agent/multi-agent-patterns/index.md), checked 2026-10-30) | **Yes.** Multi-agent graphs and subgraphs in LangGraph. ([source](https://docs.langchain.com/oss/javascript/langchain/agents), checked 2026-10-23) | **Yes.** guildAgentTool calls owner~agent; Native, Goose, and OpenClaw agents declare sub_agents in guild.yaml. ([source](https://docs.guild.ai/guide/guild-yaml.md), checked 2026-10-23) |

### Tools & integrations

| Feature | Claude Managed Agents | AWS | LangChain | Guild |
| --- | --- | --- | --- | --- |
| Managed integrations catalog | **Partial.** Built-in toolset (bash, files, web search, web fetch) and pre-built Anthropic skills (pptx, xlsx, docx, pdf). Third-party systems connect as remote MCP servers you add by URL; no managed connector catalog is documented. ([source](https://platform.claude.com/docs/en/managed-agents/onboarding.md), checked 2026-10-30) | **Partial.** AgentCore Gateway has 16 provider templates (Slack, Jira, Salesforce, ServiceNow, Zendesk, Microsoft, …) that each expose a subset of the provider's API, plus Web Search and Step Functions connectors. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-target-integrations.html), checked 2026-10-30) | **Partial.** Fleet has built-in connections (Gmail, Slack, Google Calendar, GitHub); code agents bring their own tools. ([source](https://docs.langchain.com/langsmith/fleet/mcp-framework.md), checked 2026-10-23) | **Yes.** 37 non-model integrations (GitHub, Slack, Gmail, Linear, Jira, Notion, Zendesk, …) plus 8 model publishers. ([source](https://docs.guild.ai/integrations/overview.md), checked 2026-10-23) |
| Use MCP servers as tools | **Yes.** mcp_servers on the agent plus an mcp_toolset per server; remote streamable-HTTP servers, or private servers through MCP tunnels (research preview). ([source](https://platform.claude.com/docs/en/managed-agents/mcp-connector.md), checked 2026-10-30) | **Yes.** AgentCore Gateway takes MCP servers as targets; Harness and Strands connect to remote MCP servers directly. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-target-MCPservers.html), checked 2026-10-30) | **Yes.** @langchain/mcp-adapters over stdio, streamable HTTP, or SSE. ([source](https://docs.langchain.com/oss/javascript/langchain/mcp), checked 2026-10-23) | **Yes.** Build custom integrations from MCP servers with guild integration. ([source](https://docs.guild.ai/integrations/overview.md), checked 2026-10-23) |
| Platform exposed over MCP | **No.** No MCP server exposing agents or sessions is documented. The MCP server at platform.claude.com/docs/mcp searches the docs only. ([source](https://platform.claude.com/llms.txt), checked 2026-10-30) | **Yes.** Gateway serves tools as an MCP endpoint, Runtime hosts MCP servers, and the AgentCore MCP Server lets coding assistants deploy and invoke agents. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/mcp-getting-started.html), checked 2026-10-30) | **Partial.** A Fleet Tool Server exposes custom MCP tools. ([source](https://docs.langchain.com/langsmith/fleet/mcp-framework.md), checked 2026-10-23) | **Yes.** Guild is an MCP server: https://api.guild.ai/mcp (OAuth + PKCE) or local `guild mcp`. ([source](https://docs.guild.ai/api-reference/mcp.md), checked 2026-10-23) |
| Custom APIs from OpenAPI | **No.** No OpenAPI import; use custom tools (JSON Schema) or an MCP server. ([source](https://platform.claude.com/docs/en/managed-agents/tools.md), checked 2026-10-30) | **Yes.** AgentCore Gateway turns OpenAPI schemas, Smithy models, Lambda functions, and API Gateway stages into MCP tools. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-schema-openapi.html), checked 2026-10-30) | **No.** No OpenAPI importer documented; wrap APIs as tools or MCP servers. ([source](https://docs.langchain.com/oss/javascript/langchain/mcp), checked 2026-10-23) | **Yes.** guild integration builds tools from OpenAPI specs; Porter drafts integrations from public APIs. ([source](https://docs.guild.ai/integrations/overview.md), checked 2026-10-23) |
| Secrets never reach the agent | **Yes.** Vaults hold MCP OAuth, bearer, and environment-variable credentials. Environment-variable secrets sit in the sandbox as placeholders and are substituted at egress, so the agent never sees the real value. ([source](https://platform.claude.com/docs/en/managed-agents/vaults.md), checked 2026-10-30) | **Partial.** AgentCore Gateway injects OAuth and API key credentials from the Identity token vault on outbound calls; code on Runtime that uses @requires_access_token receives the token itself. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/key-features-and-benefits.html), checked 2026-10-30) | **Partial.** Agent Auth manages OAuth per workspace; code agents handle their own secrets. ([source](https://docs.langchain.com/langsmith/fleet/essentials.md), checked 2026-10-23) | **Yes.** A credential proxy injects auth server-side; GitHub uses short-lived GitHub App tokens. ([source](https://docs.guild.ai/platform/credentials.md), checked 2026-10-23) |
| Per-operation credential policies | **Partial.** Each vault credential can be limited to allowed_hosts, and individual MCP tools can be disabled or given their own permission policy. No ALLOW/DENY rules by operation and resource on a credential. ([source](https://platform.claude.com/docs/en/managed-agents/vaults.md), checked 2026-10-30) | **Yes.** Policy in AgentCore: Cedar or Dogwood permit/forbid rules on every Gateway tool call, by principal, tool, and input values; default-deny, forbid wins. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/policy-core-concepts.html), checked 2026-10-30) | **No.** Not documented beyond workspace RBAC. ([source](https://docs.langchain.com/langsmith/rbac), checked 2026-10-23) | **Yes.** ALLOW/DENY rules by operation, resource, agent, and workspace; DENY wins. ([source](https://docs.guild.ai/platform/credential-policies.md), checked 2026-10-23) |
| Act as the invoking user | **Yes.** A vault is the set of credentials for one end user, attached per session with vault_ids, so the same agent acts with each user's own accounts. ([source](https://platform.claude.com/docs/en/managed-agents/vaults.md), checked 2026-10-30) | **Yes.** AgentCore Identity and Gateway support 3-legged OAuth and on-behalf-of token exchange carrying both the user's and the agent's identity. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/gateway-outbound-auth.html), checked 2026-10-30) | **Yes.** Fleet Assistants act with the invoking person's OAuth. ([source](https://docs.langchain.com/langsmith/fleet), checked 2026-10-23) | **Yes.** Member-scoped credentials plus identity mapping for Slack and GitHub users. ([source](https://docs.guild.ai/platform/credentials.md), checked 2026-10-23) |

### Runtime & triggers

| Feature | Claude Managed Agents | AWS | LangChain | Guild |
| --- | --- | --- | --- | --- |
| Managed hosted runtime | **Yes.** Anthropic runs the agent loop and a per-session cloud sandbox; you send events and stream results over SSE. ([source](https://platform.claude.com/docs/en/managed-agents/overview.md), checked 2026-10-30) | **Yes.** AgentCore Runtime is serverless and runs each session in its own microVM; Harness runs on top of it. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/harness-vs-runtime.html), checked 2026-10-30) | **Yes.** LangSmith Deployment Cloud (Plus plan and up). ([source](https://docs.langchain.com/langsmith/deployment), checked 2026-10-23) | **Yes.** Guild hosts and runs all agent types. ([source](https://docs.guild.ai/platform/security-architecture.md), checked 2026-10-23) |
| Self-host option | **Partial.** Self-hosted sandboxes move tool execution to your infrastructure via an environment worker; the agent loop, model calls, skills, and memory stores stay with Anthropic. ([source](https://platform.claude.com/docs/en/managed-agents/self-hosted-sandboxes.md), checked 2026-10-30) | **Partial.** AgentCore is AWS-managed only. Runtime Instances run on AWS-managed EC2 in your own AWS account; Strands agents can run anywhere without AgentCore. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-instances.html), checked 2026-10-30) | **Yes.** Hybrid, self-hosted (Enterprise), or standalone Agent Server with your own Postgres and Redis. ([source](https://docs.langchain.com/langsmith/deployment), checked 2026-10-23) | **No.** Hosted only; no self-host option documented. ([source](https://docs.guild.ai/platform/security-architecture.md), checked 2026-10-23) |
| Sandboxed, network-controlled execution | **Yes.** Each session runs in its own fresh container; environments set networking to unrestricted or limited with allowed_hosts and optional package-manager access. ([source](https://platform.claude.com/docs/en/managed-agents/environments.md), checked 2026-10-30) | **Yes.** One microVM per session; in VPC mode, Runtime, Code Interpreter, and Browser have no internet access unless you add a NAT gateway, and security groups control egress. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/agentcore-vpc.html), checked 2026-10-30) | **No.** No sandbox described in the Agent Server docs reviewed. ([source](https://docs.langchain.com/langsmith/deployment), checked 2026-10-23) | **Yes.** Containers lose network access before the agent runs; only Guild proxies can reach out. ([source](https://docs.guild.ai/platform/security-architecture.md), checked 2026-10-23) |
| Durable long-running runs | **Yes.** Sessions are stateful and persist server-side: they go idle and resume on the next message, retry transient errors automatically (rescheduling), and keep the sandbox filesystem and history. ([source](https://platform.claude.com/docs/en/managed-agents/session-operations.md), checked 2026-10-30) | **Partial.** AgentCore Runtime sessions last up to 8 hours on microVMs (14 days on Instances) and resume on a new microVM with persisted storage; no checkpoint-and-replay of the agent loop. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-lifecycle-settings.html), checked 2026-10-30) | **Yes.** Threads with a Postgres or Mongo checkpointer. ([source](https://docs.langchain.com/langsmith/deployment), checked 2026-10-23) | **Partial.** Multi-turn sessions; they can be interrupted but not resumed. ([source](https://docs.guild.ai/llms.txt), checked 2026-10-23) |
| Schedules (cron) | **Yes.** Scheduled deployments run sessions on a cron expression with a timezone, at minute granularity, plus manual runs. ([source](https://platform.claude.com/docs/en/managed-agents/scheduled-deployments.md), checked 2026-10-30) | **Via integration.** No AgentCore scheduler; an Amazon EventBridge rule or schedule calls the Runtime endpoint. ([source](https://aws.amazon.com/blogs/storage/orchestrating-multi-agent-ai-architectures-with-amazon-s3-files/), checked 2026-10-30) | **Yes.** Cron jobs (UTC), thread-bound or stateless. ([source](https://docs.langchain.com/langsmith/cron-jobs), checked 2026-10-23) | **Partial.** CLI supports --frequency CRON --cron-expression; docs list hourly–monthly presets. ([source](https://docs.guild.ai/platform/triggers.md), checked 2026-10-23) |
| Webhook, event & API triggers | **Partial.** Sessions start from the REST API or a deployment's run endpoint. Webhooks are outbound notifications (session, deployment, vault events); no built-in inbound triggers from SaaS events. ([source](https://platform.claude.com/docs/en/managed-agents/webhooks.md), checked 2026-10-30) | **Yes.** InvokeAgentRuntime and InvokeHarness APIs with IAM or OAuth auth over HTTP, MCP, A2A, AG-UI, or WebSocket; Step Functions has an InvokeHarness state. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/runtime-invoke-agent.html), checked 2026-10-30) | **Yes.** Agent Server API plus run webhooks. ([source](https://docs.langchain.com/langsmith/cron-jobs), checked 2026-10-23) | **Yes.** Event triggers (Slack mention, GitHub PR, Zendesk ticket, …) and API triggers. ([source](https://docs.guild.ai/platform/triggers.md), checked 2026-10-23) |
| Chat channels (Slack, Teams, …) | **No.** No built-in Slack or Teams channel. Anthropic's quickstarts show a chat app built with Vercel's Chat SDK whose adapters you host to reach Slack, Teams, Discord, or WhatsApp. ([source](https://platform.claude.com/docs/en/managed-agents/quickstart.md), checked 2026-10-30) | **Via integration.** No built-in channels; an AWS sample wires Slack to AgentCore Runtime with API Gateway, Lambda, and SQS. No Teams path documented. ([source](https://aws.amazon.com/blogs/machine-learning/integrating-amazon-bedrock-agentcore-with-slack/), checked 2026-10-30) | **Yes.** Fleet triggers for Slack, Gmail, and Microsoft Teams. ([source](https://docs.langchain.com/langsmith/fleet/essentials.md), checked 2026-10-23) | **Yes.** Slack event triggers; Discord, Telegram, and Twilio integrations. ([source](https://docs.guild.ai/integrations/overview.md), checked 2026-10-23) |

### Control & governance

| Feature | Claude Managed Agents | AWS | LangChain | Guild |
| --- | --- | --- | --- | --- |
| Per-tool-call approval gate | **Yes.** Permission policies per toolset or tool: always_allow, always_ask (session pauses for a tool_confirmation), or auto (server runs, denies as high-risk, or pauses). ([source](https://platform.claude.com/docs/en/managed-agents/permission-policies.md), checked 2026-10-30) | **Yes.** Strands interrupts in a BeforeToolCallEvent hook pause a tool call until your app returns an approval; written in code. Policy temporal rules can require a prior approval action. ([source](https://strandsagents.com/docs/user-guide/sdk/interrupts/index.md), checked 2026-10-30) | **Yes.** humanInTheLoopMiddleware in code; Auto/Ask per tool with an Inbox in Fleet. ([source](https://docs.langchain.com/langsmith/fleet/essentials.md), checked 2026-10-23) | **Yes.** Approval gates are set in the Guild UI, per agent and per action: when the agent reaches a gated action, the run pauses until a designated human approves. No code needed. ([source](https://www.guild.ai/product/ai-governance), checked 2026-10-23) |
| Roles & access control | **Partial.** Workspace roles (User, Limited Developer, Developer, Admin); vaults and credentials are workspace-scoped. No per-agent sharing or access rules documented. ([source](https://platform.claude.com/docs/en/manage-claude/workspaces.md), checked 2026-10-30) | **Yes.** AWS IAM identity and resource-based policies, with AgentCore condition keys, control who can create, invoke, and read each resource. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/security_iam_service-with-iam.html), checked 2026-10-30) | **Partial.** RBAC is Enterprise-only; other plans make every user Admin. ([source](https://docs.langchain.com/langsmith/rbac), checked 2026-10-23) | **Partial.** Admin and Member org roles, restricted workspaces, and scoped API keys. ([source](https://docs.guild.ai/llms.txt), checked 2026-10-23) |
| Audit log | **Partial.** Session event history is persisted and inspectable; the Compliance API Activity Feed records Console and API actions. No agent-specific immutable audit log is documented. ([source](https://platform.claude.com/docs/en/manage-claude/compliance-activity-feed.md), checked 2026-10-30) | **Yes.** AWS CloudTrail records AgentCore management and data events, including every harness and runtime invocation. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/harness-operations.html), checked 2026-10-30) | **Partial.** Enterprise-only, kept up to 400 days. ([source](https://docs.langchain.com/langsmith/audit-logs), checked 2026-10-23) | **Yes.** Immutable, CSV-exportable audit log (admins). ([source](https://docs.guild.ai/llms.txt), checked 2026-10-23) |
| Runtime guardrails & scanning | **Partial.** The auto permission policy has the server evaluate each tool call and deny high-risk ones; model safety classifiers apply. No PII or custom policy scanning documented. ([source](https://platform.claude.com/docs/en/managed-agents/permission-policies.md), checked 2026-10-30) | **Yes.** Amazon Bedrock Guardrails inside AgentCore Policy screen Gateway traffic for prompt injection, jailbreaks, PII, and harmful content. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/policy-guardrails-getting-started.html), checked 2026-10-30) | **Partial.** Middleware hooks such as humanInTheLoopMiddleware. ([source](https://docs.langchain.com/oss/javascript/langchain/agents), checked 2026-10-23) | **No.** No prompt or tool-call scanning documented; isolation comes from the network-less runtime and credential policies. ([source](https://docs.guild.ai/platform/security-architecture.md), checked 2026-10-23) |
| Spend budgets & usage limits | **Yes.** Session budgets are hard dollar caps that pause a session at its limit, also settable per deployment run; workspace spend and rate limits in the Console. ([source](https://platform.claude.com/docs/en/managed-agents/budgets.md), checked 2026-10-30) | **Partial.** Harness caps maxTokens, maxIterations, and time per invocation; Gateway rate limits cap tokens per minute by caller or model. No per-agent dollar budget in AgentCore. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/harness-operations.html), checked 2026-10-30) | **Partial.** Usage metered in LSUs and deployment compute; no per-agent budget documented. ([source](https://www.langchain.com/pricing), checked 2026-10-23) | **Yes.** Per-agent and per-workspace spend budgets, model allowlists, and daily token limits. ([source](https://docs.guild.ai/llms.txt), checked 2026-10-23) |
| Multi-provider models with fallbacks | **No.** Claude models only (Fable, Opus, Sonnet, Haiku), one per agent with per-session overrides; no other providers or automatic fallback documented. ([source](https://platform.claude.com/docs/en/managed-agents/overview.md), checked 2026-10-30) | **Yes.** Harness uses Bedrock, OpenAI, Gemini, or any LiteLLM-compatible provider and can switch mid-session; Gateway rules add ordered model fallback chains. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/release-notes.html), checked 2026-10-30) | **Partial.** provider:model strings in createAgent; fallbacks are up to your code. ([source](https://docs.langchain.com/oss/javascript/langchain/agents), checked 2026-10-23) | **Yes.** guild.yaml models is an ordered fallback list across Anthropic, OpenAI, Gemini, Meta, DeepSeek, and more; BYO keys supported. ([source](https://docs.guild.ai/guide/guild-yaml.md), checked 2026-10-23) |

### Observability

| Feature | Claude Managed Agents | AWS | LangChain | Guild |
| --- | --- | --- | --- | --- |
| Run traces | **Yes.** Console session viewer: timeline, transcript with thinking and tool calls, raw events, per-tool stats, threads; OpenTelemetry export in beta. ([source](https://platform.claude.com/docs/en/managed-agents/session-observability.md), checked 2026-10-30) | **Yes.** AgentCore Observability: OpenTelemetry traces of model calls, tool calls, and memory operations in Amazon CloudWatch. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/harness-operations.html), checked 2026-10-30) | **Yes.** LangSmith tracing (14-day base, 180-day extended retention). ([source](https://www.langchain.com/pricing), checked 2026-10-23) | **Yes.** Session event log of LLM calls, tool calls, sub-tasks, and container logs; also via the API. ([source](https://docs.guild.ai/llms.txt), checked 2026-10-23) |
| Evaluations | **Partial.** Outcomes give a session a rubric and a separate grader that iterates the agent until it passes. No dataset-based eval runs for agents documented. ([source](https://platform.claude.com/docs/en/managed-agents/define-outcomes.md), checked 2026-10-30) | **Yes.** AgentCore Evaluations (GA March 2026): built-in and custom evaluators, online, on-demand, and batch runs, plus A/B tests. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/evaluations.html), checked 2026-10-30) | **Yes.** Datasets, evaluators, and Engine for recurring failures. ([source](https://www.langchain.com/pricing), checked 2026-10-23) | **Partial.** Evaluations with samples, checks, and llmJudge; enabled per account, no CLI command yet. ([source](https://docs.guild.ai/platform/evaluations.md), checked 2026-10-23) |
| Cost & usage tracking | **Yes.** Per-session usage and list cost on the session object and in the Console; usage grouped by agent on the Usage page and in the usage APIs. ([source](https://platform.claude.com/docs/en/managed-agents/session-observability.md), checked 2026-10-30) | **Partial.** Token usage in invocation metadata and traces; billed cost via AWS Cost Explorer with harness tags as cost allocation tags. AWS says observability is not a billing report. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/harness-operations.html), checked 2026-10-30) | **Partial.** Usage visible per LSU and deployment; trace-level cost depends on setup. ([source](https://www.langchain.com/pricing), checked 2026-10-23) | **Yes.** Insights and Spend Explorer. ([source](https://docs.guild.ai/llms.txt), checked 2026-10-23) |

### Ecosystem & business

| Feature | Claude Managed Agents | AWS | LangChain | Guild |
| --- | --- | --- | --- | --- |
| Public agent hub (install & fork) | **No.** No public agent hub. The Console offers quickstart templates, and Anthropic publishes example skills on GitHub. ([source](https://platform.claude.com/docs/en/managed-agents/onboarding.md), checked 2026-10-30) | **Partial.** AWS Marketplace lists third-party agents and tools you buy and deploy to AgentCore; AWS Agent Registry is a private catalog. No public install-and-fork hub. ([source](https://aws.amazon.com/about-aws/whats-new/2025/07/ai-agents-tools-aws-marketplace/), checked 2026-10-30) | **Partial.** Fleet templates and the Prompt Hub; no public agent marketplace. ([source](https://docs.langchain.com/langsmith/fleet), checked 2026-10-23) | **Yes.** Agent Hub: Draft, Internal, or Public; owner~name with semver; install and fork (guild agent init --fork). ([source](https://docs.guild.ai/platform/publish-to-agent-hub.md), checked 2026-10-23) |
| Reusable skills | **Yes.** Pre-built Anthropic skills and custom skills uploaded to the workspace, or loaded from a mounted GitHub repository's .claude/skills. ([source](https://platform.claude.com/docs/en/managed-agents/skills.md), checked 2026-10-30) | **Yes.** Harness loads SKILL.md skills from the AWS skills catalog, Git, Amazon S3, or the session filesystem; Registry catalogs skills. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/harness-skills.html), checked 2026-10-30) | **Partial.** Deep Agents and the Prompt Hub cover reusable instructions. ([source](https://docs.langchain.com/langsmith/fleet), checked 2026-10-23) | **Yes.** Hub skills, `guild skill`, and skills inside Goose and OpenClaw agents. ([source](https://docs.guild.ai/cli/commands.md), checked 2026-10-23) |
| Full lifecycle CLI | **Yes.** The ant CLI: ant apply syncs agents, environments, skills, memory stores, deployments, and vaults from files; ant beta:sessions connect follows a live session; ant beta:worker runs self-hosted workers. ([source](https://platform.claude.com/docs/en/cli-sdks-libraries/cli/apply.md), checked 2026-10-30) | **Yes.** AgentCore CLI v1 (@aws/agentcore): create, add, dev, deploy, invoke, logs, traces, eval, plus harness, runtime, gateway, and memory commands. ([source](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/release-notes.html), checked 2026-10-30) | **Yes.** langgraph dev, build, up, deploy, plus deploy logs and revisions. ([source](https://docs.langchain.com/langsmith/cli), checked 2026-10-23) | **Yes.** 17 top-level commands: init, test, save --publish, workspace chat, trigger, integration, skill, session, and more. ([source](https://docs.guild.ai/cli/commands.md), checked 2026-10-23) |
| Open-source core | **No.** The platform is proprietary. The client SDKs and ant CLI are open source (Apache 2.0 since 2026-10-21). ([source](https://platform.claude.com/docs/en/release-notes/overview.md), checked 2026-10-30) | **Partial.** Strands Agents, the AgentCore CLI, and the AgentCore Python SDK are Apache-2.0; the AgentCore services themselves are proprietary. ([source](https://github.com/aws/agentcore-cli), checked 2026-10-30) | **Yes.** LangChain and LangGraph are open source; LangSmith is proprietary. ([source](https://www.langchain.com/pricing), checked 2026-10-23) | **No.** Proprietary platform; the CLI is published on npm. ([source](https://docs.guild.ai/llms.txt), checked 2026-10-23) |
| Public pricing | **Yes.** Tokens at standard model rates plus $0.08 per session-hour while a session is running; web search $10 per 1,000. ([source](https://platform.claude.com/docs/en/about-claude/pricing.md), checked 2026-10-30) | **Yes.** Per-service rates published (e.g. Runtime $0.0895 per vCPU-hour, Gateway $0.005 per 1,000 invocations); no separate Harness charge. ([source](https://aws.amazon.com/bedrock/agentcore/pricing/), checked 2026-10-30) | **Yes.** Developer $0, Plus $39 per seat per month, Enterprise custom. Fleet free-tier access is unclear. ([source](https://www.langchain.com/pricing), checked 2026-10-23) | **Partial.** No public price list: plans are priced with the Guild team; model usage is metered at list price. ([source](https://docs.guild.ai/reference/pricing.md), checked 2026-10-23) |

## More

- [Every platform, every cell (llms.txt)](https://www.guild.ai/compare/llms.txt)
- [Migrate to Guild](https://www.guild.ai/migrate)
- [Book a demo](https://www.guild.ai/talk-to-sales)