AI agents #001 VCG
Tools for building and managing AI agents
Compare 34 tools by the work you want to delegate. Find a starting point and see where Claude, OpenAI and Gemini fit.
AI AGENTS MARKET MAP
How to read and cite this map
The four zones group user needs. Seven categories distinguish products by the unit of work they handle. The bottom band groups building blocks for custom systems. Position does not score quality, autonomy or market share. This editorial selection can expand.
Documented capabilities link to primary sources and their review dates. Unverified fields are not incompatibilities. API, runtime and MCP client are different connection paths; check the conditions.
Source: Vibe Coding Gallery, AI Agents Market Map, 16 November 2026 edition. vibecoding.gallery
Explore the tools
Open a profile to see what you can delegate, what to check and the source documentation.
Filters show documented capabilities. Check the conditions in each profile.
Personal assistants
Self-hosted assistants
Coding agents
Workflows and agent builders
Business operations
Team coordination
Frameworks and runtimes
No tools match your filters.
Tool details and evidence
Check the connections, deployment and controls we could document. Suggested uses are examples; each capability cites its source and review date.
InstinctPersonal assistants
Suggested use. Track a Notion list you still edit outside chat.
What to check Can the assistant edit the right page and reflect its current state?
- Channels and integrations
-
The provider describes messaging and calls. The reviewed page does not establish a WhatsApp Business API connector.
Vendor statement · instinct.com ↗
Not verified. How you build · Deployment and execution · Models and authentication · Human approval · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗DotsPersonal assistants
Suggested use. Prepare work that continues between conversations.
What to check Which apps can it open, and does it need access to your computer?
- Deployment and execution
-
An agent inside ChatGPT with its own computer and the apps you connect.
Documentation reviewed · help.openai.com ↗ - State and recovery
-
Can continue work between conversations and start from ChatGPT memory. This does not document transactional recovery guarantees.
Documentation reviewed · help.openai.com ↗ - Availability
-
The reviewed help limits the Pro rollout to adults outside the EEA, Switzerland and UK. Business Premium depends on workspace eligibility.
Documentation reviewed · help.openai.com ↗
Not verified. How you build · Models and authentication · Human approval · Channels and integrations · Billing. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Claude DuetPersonal assistants
Suggested use. Review commitments and conversations in one thread.
What to check What sources feed reminders, and how do you review an action?
- Human approval
-
The announcement describes checking with the user before acting on their behalf. It is not a specification for programmable approval of every tool.
Vendor statement · claude.com ↗ - Availability
-
Research preview with a waitlist, initially for US Pro, Max and Ultra subscribers.
Documentation reviewed · claude.com ↗
Not verified. How you build · Deployment and execution · Models and authentication · Channels and integrations · State and recovery · Billing. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Grok BotPersonal assistants
Suggested use. Repeat a task on a schedule.
What to check What can trigger the routine, and how do you correct a run?
- Deployment and execution
-
Bots share a persistent cloud computer with a browser, files and terminal.
Documentation reviewed · docs.x.ai ↗ - Channels and integrations
-
Documentation describes desktop, mobile and Slack access. Connectors and permissions are configured for the task.
Documentation reviewed · docs.x.ai ↗
Not verified. How you build · Models and authentication · Human approval · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Meta MusePersonal assistants
Suggested use. Coordinate follow-ups in the app or on WhatsApp where available.
What to check Check regional availability and which actions require permission.
- Deployment and execution
-
According to Meta, Muse runs the agent and its data in a dedicated Muse Secure VM.
Vendor statement · about.fb.com ↗ - Channels and integrations
-
Meta describes personal use in the Muse app or WhatsApp. This does not establish a platform for deploying business bots on the WhatsApp Business API.
Vendor statement · about.fb.com ↗
Not verified. How you build · Models and authentication · Human approval · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗OpenClawSelf-hosted assistants
Suggested use. Put an assistant in the channels you choose.
What to check Who manages the server, credentials and access?
- Deployment and execution
-
The Gateway can run on your computer or your own server. You manage installation, updates, backups and credentials.
Documentation reviewed · docs.openclaw.ai ↗ - Models and authentication
-
Documents Anthropic, OpenAI and Gemini API access plus Claude CLI, Codex and Gemini CLI runtime routes. Selecting a runtime alone does not determine billing; Gemini CLI uses the documented AI Studio key path.
- Channels and integrations
-
Its WhatsApp channel uses a linked WhatsApp Web session through Baileys. This is a different connection from the WhatsApp Business Cloud API.
Documentation reviewed · docs.openclaw.ai ↗ - Billing
-
Anthropic API access is billed by usage. The Claude CLI route reuses Claude Code login on the same host; a selected API account still uses API billing. Check both account and route before assuming your subscription applies.
Documentation reviewed · docs.openclaw.ai ↗
Not verified. How you build · Human approval · State and recovery · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Hermes AgentSelf-hosted assistants
Suggested use. Configure tools and skills in a deployable agent.
What to check How are instructions stored and access to tools limited?
- Deployment and execution
-
The repository documents local execution, Docker and SSH, plus sandbox backends. Your chosen infrastructure determines where it runs.
Documentation reviewed · github.com ↗
Not verified. How you build · Models and authentication · Human approval · Channels and integrations · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗n8nWorkflows and agent builders
Suggested use. Process a request and require approval before a tool call.
What to check Where does the workflow pause, and how do you prevent duplicate actions?
- Deployment and execution
-
You can deploy n8n on your infrastructure. n8n Cloud is also available; editions and features differ.
- Models and authentication
-
Anthropic, OpenAI and Gemini nodes use provider credentials or n8n Cloud Gateway credits. They do not authenticate with a chat subscription.
- Human approval
-
You can require approval before a specific AI Agent tool runs. Configure the protected tools and review channel.
Documentation reviewed · docs.n8n.io ↗ - Channels and integrations
-
Includes a WhatsApp Business Cloud node. Sending and waiting for a reply require WhatsApp Business credentials.
Documentation reviewed · docs.n8n.io ↗ - Billing
-
Separate n8n hosting or plan costs from model usage. The reviewed nodes support your own key or Cloud gateway credits.
Not verified. How you build · State and recovery · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Zapier AgentsWorkflows and agent builders
Suggested use. Respond to events and act in connected apps.
What to check Which integration supports the action and what identity will it use?
- How you build
-
Configure agents with company knowledge and actions across connected Zapier applications.
Vendor statement · zapier.com ↗
Not verified. Deployment and execution · Models and authentication · Human approval · Channels and integrations · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗GumloopWorkflows and agent builders
Suggested use. Chain document processing with other apps.
What to check Where are retries, inputs and execution records visible?
- Models and authentication
-
The documented MCP connection lets Claude use Gumloop. This verifies a client connection, not the internal model catalog.
Documentation reviewed · docs.gumloop.com ↗ - Channels and integrations
-
Documentation offers REST API execution and an MCP server for using Gumloop from Claude.
Documentation reviewed · docs.gumloop.com ↗
Not verified. How you build · Deployment and execution · Human approval · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗LindyWorkflows and agent builders
Suggested use. Draft email or meeting follow-ups.
What to check Can you check a draft before it uses your account?
- Human approval
-
The provider describes initial approval for externally impactful actions and later delegation by job. Scope depends on what you authorize.
Vendor statement · lindy.ai ↗ - Channels and integrations
-
The provider lists Slack, email, messaging and MCP among its work channels. Check permissions on connected accounts.
Vendor statement · lindy.ai ↗
Not verified. How you build · Deployment and execution · Models and authentication · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Relevance AIWorkflows and agent builders
Suggested use. Split work among agents with tools and escalations.
What to check Where does a handoff pause for human input?
- How you build
-
A low-code/no-code visual builder for agents and teams; documentation separates agents, tools, knowledge and workforces.
Documentation reviewed · relevanceai.com ↗
Not verified. Deployment and execution · Models and authentication · Human approval · Channels and integrations · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗DifyWorkflows and agent builders
Suggested use. Build a custom assistant around your own data.
What to check How do you configure the model, data and publication?
- Deployment and execution
-
Dify documents Docker Compose deployment and Dify Cloud. Self-hosting includes services you need to operate.
- Models and authentication
-
Anthropic, OpenAI and Gemini plugins use API credentials. Models and credentials are configured in the workspace.
Documentation reviewed · marketplace.dify.ai [1] ↗marketplace.dify.ai [2] ↗marketplace.dify.ai [3] ↗docs.dify.ai [4] ↗ - Billing
-
Dify Cloud offers AI credits; your own keys bill to the provider account and are shared across the workspace.
Documentation reviewed · docs.dify.ai ↗
Not verified. How you build · Human approval · Channels and integrations · State and recovery · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Claude Code / Agent TeamsCoding agents
Suggested use. Separate implementation and review in a coding environment.
What to check Can you see the final changes and what each session contributed?
- Deployment and execution
-
Claude Code offers terminal, IDE, desktop and web surfaces. The chosen surface determines where work runs.
Documentation reviewed · code.claude.com ↗ - Models and authentication
-
Claude Code is Anthropic’s runtime. Documentation distinguishes Claude accounts, Anthropic Console and third-party providers by surface.
Documentation reviewed · code.claude.com ↗ - Billing
-
Most surfaces require a Claude subscription or Anthropic Console account. Verify the route used by any tool invoking Claude Code.
Documentation reviewed · code.claude.com ↗
Not verified. How you build · Human approval · Channels and integrations · State and recovery · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗CodexCoding agents
Suggested use. Prepare a change or review a repository.
What to check Where does the work run, and what permissions, tests and diff can you inspect?
- Deployment and execution
-
Codex CLI works with the repository and tools installed on your machine. This entry does not assume CLI and cloud tasks share the same environment.
Documentation reviewed · learn.chatgpt.com ↗ - Models and authentication
-
The CLI allows model, effort and permission selection. Sign in with ChatGPT or another available authentication method.
Documentation reviewed · learn.chatgpt.com ↗
Not verified. How you build · Human approval · Channels and integrations · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Cursor Cloud AgentsCoding agents
Suggested use. Delegate changes in a remote environment, including work across repositories.
What to check Which repositories, secrets and tests are available to the agent?
- Deployment and execution
-
Runs agents in isolated cloud VMs. Your personal computer can disconnect; repositories, dependencies and access must be prepared.
Documentation reviewed · cursor.com ↗ - Channels and integrations
-
Supports MCP servers and triggers from Slack or repositories. Source-control connections require account configuration.
Documentation reviewed · cursor.com ↗
Not verified. How you build · Models and authentication · Human approval · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗DevinCoding agents
Suggested use. Delegate a development task in an agent session.
What to check Which evidence can you review before accepting the change?
- Deployment and execution
-
Documentation distinguishes cloud Devin from Devin for Terminal; local work can be handed off to the cloud.
Documentation reviewed · docs.devin.ai ↗
Not verified. How you build · Models and authentication · Human approval · Channels and integrations · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗GitHub Copilot cloud agentCoding agents
Suggested use. Assign code changes from your repository.
What to check Which branch is affected, and who can approve the pull request?
- Channels and integrations
-
Can be started from GitHub, API, CLI or the GitHub MCP server. Work is delivered for review through pull requests.
Documentation reviewed · docs.github.com ↗
Not verified. How you build · Deployment and execution · Models and authentication · Human approval · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗PaperclipTeam coordination
Suggested use. Manage agents, dependencies, reviews and usage.
What to check What does the board handle, and what depends on the adapter or integration?
- Deployment and execution
-
Documentation distinguishes self-hosted installations and Paperclip Cloud. Local adapters need an environment with the runtime installed.
Documentation reviewed · docs.paperclip.ing ↗ - Models and authentication
-
Connects Claude Code, Codex or Gemini CLI through adapters. New agents may use Paperclip Runner; the Legacy adapter is a separate path.
Documentation reviewed · docs.paperclip.ing ↗ - Human approval
-
Coding adapters start in automatic mode by default. Company approvals and runtime permissions are separate controls.
Documentation reviewed · docs.paperclip.ing ↗ - Billing
-
Check the chosen adapter and edition credentials. The Legacy guide describes API keys; it does not establish that any chat subscription works.
Documentation reviewed · docs.paperclip.ing ↗
Not verified. How you build · Channels and integrations · State and recovery · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗MulticaTeam coordination
Suggested use. Assign work and retrieve the deliverable from the board.
What to check Which resources and context reach the agent during each run?
- Deployment and execution
-
A local runtime executes tools on the connected computer. Their code and credentials remain on that machine.
Documentation reviewed · multica.ai ↗ - Models and authentication
-
Detects Claude Code and Codex CLI among other tools. They must be installed and work independently before Multica invokes them.
Documentation reviewed · multica.ai ↗ - Billing
-
Invoking a CLI does not include its account, subscription or quota. Authentication and usage depend on the local tool.
Documentation reviewed · multica.ai ↗
Not verified. How you build · Human approval · Channels and integrations · State and recovery · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗OrcaTeam coordination
Suggested use. Work in parallel sessions with separate worktrees.
What to check Where are changes visible, and which working tree will receive them?
- Deployment and execution
-
Offers local, SSH, Orca server and per-workspace VM execution on infrastructure or cloud accounts you control. It does not sell managed VPS hosting.
Documentation reviewed · orca.dev ↗
Not verified. How you build · Models and authentication · Human approval · Channels and integrations · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗SymphonyTeam coordination
Suggested use. Turn work from an issue tracker into isolated agent runs.
What to check What is configured for dispatch and how is execution scoped?
- How you build
-
You can implement the specification or run the Elixir reference. Its example coordinates Linear work with implementation agents.
Documentation reviewed · github.com ↗ - Availability
-
The repository describes an engineering preview for trusted environments with an experimental reference implementation.
Documentation reviewed · github.com ↗
Not verified. Deployment and execution · Models and authentication · Human approval · Channels and integrations · State and recovery · Billing. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗CofounderBusiness operations
Suggested use. Connect an agent to your CRM, email, deployment or payment tools.
What to check What identity does it use, and which actions require approval?
- Deployment and execution
-
The MCP endpoint is hosted. Client execution can use a Cofounder sandbox or a desktop terminal, depending on the surface.
Documentation reviewed · docs.cofounder.co ↗ - Models and authentication
-
Claude, Claude Code, ChatGPT or Codex act as Cofounder clients through MCP or CLI. This connects business tools rather than selecting an internal Cofounder model.
Documentation reviewed · docs.cofounder.co ↗ - Billing
-
After initial setup, using your own agent requires a Cofounder company subscription. Also check the client plan and access.
Documentation reviewed · docs.cofounder.co ↗
Not verified. How you build · Human approval · Channels and integrations · State and recovery · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗ChatGPT Workspace AgentsBusiness operations
Suggested use. Create, share or schedule an agent for the team.
What to check Under whose credentials will it use connected apps?
- Channels and integrations
-
Agents can be shared in a workspace, used in Slack, scheduled or triggered through an API, subject to configuration and permissions.
Documentation reviewed · help.openai.com ↗ - Availability
-
Availability depends on Business and Enterprise workspace eligibility and permissions. Enterprise requires admin enablement.
Documentation reviewed · help.openai.com ↗
Not verified. How you build · Deployment and execution · Models and authentication · Human approval · State and recovery · Billing. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Microsoft Copilot StudioBusiness operations
Suggested use. Combine agents, flows and connectors in channels such as Teams.
What to check Which actions use employee identity and admin permissions?
- How you build
-
A graphical low-code studio for building agents and workflows with prebuilt or custom connectors.
Documentation reviewed · learn.microsoft.com ↗ - Channels and integrations
-
Documentation describes publishing to Teams, Microsoft Copilot, websites and mobile apps, depending on channel configuration.
Documentation reviewed · learn.microsoft.com ↗
Not verified. Deployment and execution · Models and authentication · Human approval · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Salesforce AgentforceBusiness operations
Suggested use. Work on cases and processes using Salesforce data.
What to check Can the agent read and write only the records you intend?
- Channels and integrations
-
The provider offers sales and service agents with prebuilt actions and Salesforce context. Verify the edition and CRM permissions.
Vendor statement · salesforce.com ↗
Not verified. How you build · Deployment and execution · Models and authentication · Human approval · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Google Gemini EnterpriseBusiness operations
Suggested use. Find information and delegate actions in an enterprise environment.
What to check Which sources are connected and which actions are allowed?
- Deployment and execution
-
Google describes long-running jobs on a secure Google Cloud machine with the user’s existing permissions.
Vendor statement · cloud.google.com ↗
Not verified. How you build · Models and authentication · Human approval · Channels and integrations · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Intercom FinBusiness operations
Suggested use. Resolve questions in the customer support channel.
What to check When does the agent hand the conversation to a person?
- Models and authentication
-
The provider states that Fin uses its Apex models. The page does not establish user-selectable Claude, OpenAI or Gemini models.
Vendor statement · fin.ai ↗
Not verified. How you build · Deployment and execution · Human approval · Channels and integrations · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗LangGraphFrameworks and runtimes
Suggested use. Program an agent that can pause and resume.
What to check What state can you persist and recover after a failure?
- How you build
-
Control is defined in a programmed graph. Your application implements the review interface and each node’s effects.
Documentation reviewed · docs.langchain.com ↗ - Human approval
-
You can program interrupt() and resume with a human decision. A checkpointer, thread ID and explicitly designed approval point are required.
Documentation reviewed · docs.langchain.com ↗ - State and recovery
-
Checkpointers save state by thread. Production recovery needs persistent storage; an in-memory saver is not sufficient.
Not verified. Deployment and execution · Models and authentication · Channels and integrations · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗CrewAIFrameworks and runtimes
Suggested use. Define roles, tasks and coordination.
What to check What happens if the next agent cannot complete a handoff?
- How you build
-
The framework installs in your environment and organizes agents, crews and flows. Documentation separates local development from Enterprise deployment.
Documentation reviewed · docs.crewai.com ↗ - State and recovery
-
Flows document state management, persistence and resumption. Configure the flow and storage for the specific use case.
Documentation reviewed · docs.crewai.com ↗
Not verified. Deployment and execution · Models and authentication · Human approval · Channels and integrations · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Google ADKFrameworks and runtimes
Suggested use. Implement the logic of agents and their workflows.
What to check Which deployment and observability options match your environment?
- How you build
-
An agent development framework with programmed control over tools and workflows. You need to build and deploy the application.
Documentation reviewed · adk.dev ↗ - Models and authentication
-
The official example configures an agent with a Gemini model. Equivalent support across every provider or language is not inferred.
Documentation reviewed · adk.dev ↗
Not verified. Deployment and execution · Human approval · Channels and integrations · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗OpenAI Agents SDKFrameworks and runtimes
Suggested use. Implement and observe delegations in code.
What to check Which tool calls and handoffs are visible in traces?
- How you build
-
A Python library with agents, tools, handoffs, guardrails and tracing. The application defines its logic and environment.
Documentation reviewed · openai.github.io ↗ - Models and authentication
-
Uses the Responses API by default with OpenAI models. Other providers require checking their specific configuration.
Documentation reviewed · openai.github.io ↗ - State and recovery
-
Includes a session layer for maintaining context. Specific persistence and recovery depend on the chosen implementation.
Documentation reviewed · openai.github.io ↗
Not verified. Deployment and execution · Human approval · Channels and integrations · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Claude Managed AgentsFrameworks and runtimes
Suggested use. Operate asynchronous work and sessions that continue later.
What to check Which permissions, isolation and persistence are handled for you?
- Deployment and execution
-
Documents Anthropic-managed sandboxes and sandboxes on your infrastructure. Self-hosting a sandbox does not by itself mean self-hosting the entire service.
Documentation reviewed · platform.claude.com ↗ - State and recovery
-
Maintains event history and sessions with files and conversation state across interactions.
Documentation reviewed · platform.claude.com ↗
Not verified. How you build · Models and authentication · Human approval · Channels and integrations · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗AWS Bedrock AgentCoreFrameworks and runtimes
Suggested use. Host and run your agent using managed components.
What to check Which components are managed, and which settings do you still own?
- Deployment and execution
-
An AWS platform for running and operating agents. It supports different frameworks; model choice and access are configured separately.
Vendor statement · aws.amazon.com ↗
Not verified. How you build · Models and authentication · Human approval · Channels and integrations · State and recovery · Billing · Availability. Lack of verification does not mean the tool does not support it.
Tool documentation ↗ Link to this tool ↗Cite or share the map
Include the edition and link to VCG so readers can find sources and updates.
A customer sends you a question. Do you want an agent to remind you to reply, draft an answer, or update the CRM? Each job calls for a different tool. If you want to delegate a bug fix, the deliverable changes again: you need to see the code change and decide whether to accept it.
Choose by what you want to have done. This guide compares 34 tools for building and managing AI agents. Start with the job, then use your shortlist to check which models the tools support, where they run, and what you can review.
What kind of agent do you need?
Pick the job closest to yours. Each option leads to an example and a way to test it.
- 01 / Request
taskReproduce the bug
- 02 / Execution
agent.run()Model, tools and permissions
- 03 / Deliverable
pull_requestChange ready for review
- Personal assistantsDelegate reminders and follow-upsDeliverable A reminder about something pendingInstinct · Dots · Claude Duet · Grok Bot · Meta Muse
- Assistants on your own infrastructureChoose where the assistant runsDeliverable An assistant in an environment you controlOpenClaw · Hermes Agent
- Workflows and agent buildersDesign the process or add it to your productDeliverable A recorded request or a completed workflown8n · Dify · LangGraph
- Coding agentsDelegate implementation and reviewDeliverable A diff or pull request ready for reviewClaude Code · Codex · Agent Teams
- Team coordinationTrack the work across ownersDeliverable A deliverable with an owner and statusPaperclip · Multica
- Business operationsConnect an agent to company accountsDeliverable An action recorded in the CRMCofounder
How AI agent tools work
An agent system combines several parts. The model decides what to do with the information it receives. The runtime opens files, calls tools and runs commands. The orchestrator assigns the task, decides when to start it and links the result to the original request. Claude is a family of models, Claude Code is the program doing the work, and Paperclip can start that program when a task arrives.
| What you want to delegate | How it starts | What you can check afterwards |
|---|---|---|
| A reminder about something pending | You message an assistant, or a connected app sends it a signal. | A reminder or draft you can review in its channel. |
| Draft a reply for every new request | An n8n trigger starts the workflow. A node calls the model and the workflow controls where the result goes. | A workflow execution with inputs and outputs for each step. |
| Implement a change that requires review | You assign a task with a repository and context. The board dispatches it to an agent that can run tools. | A deliverable, such as a diff or document, linked to the task. |
| Work with business data or actions | An agent uses tools exposed by a connection such as Cofounder. | A read or action on the connected system, subject to its permissions. |
Where to start with each type of agent
Personal assistants
You get a message saying “remind me to review the budget on Thursday.” The assistant’s job is to bring it back at the right time. Instinct, Dots, Claude Duet, Grok Bot and Meta Muse are starting points if the work begins in conversations or routines. Try two reminders, change one in the source app and see what comes back. Check whether a reminder is enough or whether you need the assistant to act for you.
Assistants on your own infrastructure
You want something similar, with control over the machine, model or tools. With OpenClaw or Hermes Agent, the person operating the assistant takes on part of the setup. Start with a test folder and a read-only tool. Before connecting real accounts, check which folders it can see, which permissions it inherits, and what happens when you turn off the machine. OpenClaw · Hermes Agent.
Workflows and agent builders
Every time a form arrives, you want to classify it, draft a reply and save it for review. n8n, Gumloop or Zapier Agents fit when you know what starts the process and where the result should go. Look at Dify if you need your own interface, Relevance AI if you want to configure a team of agents, or LangGraph if you are implementing a process with state in code. Test what happens when a field is missing or the same request arrives twice. n8n · Dify · LangGraph.
Coding agents
You want to fix a specific bug in your repository. Start with Claude Code or Codex: give it the failure, a way to reproduce it and a criterion for accepting the change. Ask for the diff and the tests it ran. Agent Teams fits work you can split, such as having one session investigate the cause while another reproduces the issue. Someone still needs to combine the findings and decide which change goes in. Agent Teams.
Team coordination
The task moves from research to implementation and then to review. You want to see who owns it, what context they received and what happens if someone requests a change. That’s when it makes sense to compare Paperclip and Multica. To test them, have a reviewer send back a deliverable with a change request and check that the right agent receives the context it needs.
Business operations
You want an agent to look up an account in your CRM and draft the next step. Cofounder is one way to connect an agent you already use. If sales or support already runs inside another platform, check what its own agents can do there. Start with a read and a result you can review. Check permissions, handoffs and records before enabling sending or data changes.
Which tools support Claude, OpenAI and Gemini?
| Platform | Claude / Anthropic | OpenAI | Google Gemini | What you configure |
|---|---|---|---|---|
| n8n | Anthropic Chat Model node | OpenAI Chat Model node | Google Gemini Chat Model node | Provider credentials or gateway credits where available. Connect the model to the agent node and add integrations to the workflow. |
| Dify | Anthropic plugin | OpenAI plugin | Gemini plugin | Install a model provider. Use Dify Cloud credits or a workspace API key, depending on the model. |
| Relevance AI | Claude models in its catalog | OpenAI models | Gemini models | Choose a model in the builder. Check the catalog and credits available to your workspace. |
| OpenClaw | Anthropic provider | OpenAI provider | Google provider | Configure the provider and its authentication method. Operate the assistant and its integrations in your chosen environment. |
| Paperclip | Claude Code adapter | Codex adapter | Gemini CLI adapter, depending on environment | Configure the adapter and its credentials. Local runs execute on the configured machine. Check your edition’s options. |
| Multica | Claude Code on the connected computer | Codex CLI on the connected computer | Through OpenCode configured with Google | The daemon detects the supported program. The model and authentication depend on that program and the connected computer. |
| Cofounder | Connect from Claude or Claude Code | Connect from ChatGPT or Codex | Check whether your Gemini client supports the connection | Connect your agent to Cofounder’s tools via MCP or its CLI. Availability depends on the client and its setup. |
Table sources: n8n: Anthropic · OpenAI · Gemini · Dify: setup · Anthropic · OpenAI · Gemini · Relevance AI · OpenClaw · Paperclip · Multica · OpenCode · Cofounder.
If you already pay for Claude, check what authenticates the run. n8n’s Anthropic node uses its own credentials or compatible credits. Multica can detect an existing, signed-in Claude Code install on the connected computer. In Paperclip, check your chosen adapter and the credentials it requires. For Cofounder, check both access from your MCP or CLI client and the permissions for the business tools you intend to use.
Paperclip vs. Multica. What happens when you assign work?
Paperclip organizes agents around roles, tasks and reporting relationships. A trigger, such as an assignment or a scheduled heartbeat, wakes the agent. The adapter launches the runtime with the task context and sends the output back to the run log. In our setup, a lead delegated work and was woken up when the deliverable arrived. That pattern works when you want someone to own the review and decide what happens next. Paperclip architecture · Adapters.
Multica organizes work in workspace issues. With a local runtime, a background process detects installed tools, picks up the task and runs the chosen program on the connected computer. The workspace retains comments and run records. The local program’s credentials stay on that computer. If your laptop is the execution host, check what happens when it goes offline and which tools the agent can access. Daemon and runtimes.
| Decision | Paperclip | Multica |
|---|---|---|
| Where work starts | An organization with roles, tasks and triggers. The adapter determines how to launch the program. | An issue, agent and assigned runtime. The local daemon picks up work for an installed program. |
| What the operator configures | Adapter, execution environment, credentials and permissions. If using MCP, check how each agent receives its configuration. | Connected machine, authenticated program and access to the repository and tools. A local runtime is private by default. |
| What I would ask to see in a demo | A task changes owner, the lead wakes up and the demo shows who accepts the result. | A reviewer finds a deliverable and requests a traceable revision in the same issue. |
We built our own review panel on top of Multica
We wanted reviewers to see what changed, which version they are commenting on and exactly what they are asking to fix in the same view. We built a local interface that reads the deliverable from Multica, shows the document and its changes, and lets the reviewer prepare a request before sending it to the agent.
- Deliverable v1Read and request a change
- Multica issueThe agent receives the comment
- Deliverable v2Review changes and the previous version
Diagram of the workflow tested with fictional data.
When the reviewer sends the request, our panel creates a comment on the Multica issue and mentions the agent. The interface distinguishes feedback recorded, run started and new version received. It displays a new deliverable once it can connect the feedback comment to the expected agent’s run and the preceding version. We tried two complete feedback rounds on a fictional case and retained three versions. The first revision described part of the interface incorrectly. We requested a correction from the same panel and received a third version with the fix.
Multica handles issues, comments and run records. Our panel defines what the reviewer needs to see and when a deliverable counts as a response to their request. To evaluate a similar workflow, try clicking send twice, submitting feedback on a stale version and following a completed run to its exact deliverable.
Cofounder. What it adds if you already use Claude Code
Cofounder lets an agent you already use work with company tools, such as reading company context, operating on tasks or looking up CRM contacts. The MCP or CLI connection authenticates to your Cofounder account and lets the agent call the available operations. Where the agent runs depends on the client. To move from “give me a strategy” to “look up this CRM record and draft a follow-up,” check what actions the connection authorizes and where you can review the outcome. Connect an agent · How your agent works.

Map of 34 tools for building and managing AI agents
Find the unit of work you want to delegate. The download includes a task, what to verify and a source for every row. Download the CSV matrix in English.
| Tool | Unit of work | Category |
|---|---|---|
| Instinct | Personal follow-up | Personal assistants |
| Dots | Personal project | Personal assistants |
| Claude Duet | Workday commitments | Personal assistants |
| Grok Bot | Assignment or routine | Personal assistants |
| Meta Muse | Errand or daily plan | Personal assistants |
| OpenClaw | Self-hosted assistant | Self-hosted assistants |
| Hermes Agent | Configurable assistant | Self-hosted assistants |
| n8n | Workflow run | Workflows and agent builders |
| Zapier Agents | Cross-app flow | Workflows and agent builders |
| Gumloop | Visual workflow | Workflows and agent builders |
| Lindy | Office task | Workflows and agent builders |
| Relevance AI | Configurable agent or team | Workflows and agent builders |
| Dify | Application or chatflow | Workflows and agent builders |
| Claude Code / Agent Teams | Session or group of sessions | Coding agents |
| Codex | Development task | Coding agents |
| Cursor Cloud Agents | Task in a VM | Coding agents |
| Devin | Engineering assignment | Coding agents |
| GitHub Copilot cloud agent | Issue or pull request | Coding agents |
| Paperclip | Task with an owner | Team coordination |
| Multica | Issue in a workspace | Team coordination |
| Orca | Workspace or worktree | Team coordination |
| Symphony | Engineering issue | Team coordination |
| Cofounder | Business operation | Business operations |
| ChatGPT Workspace Agents | Shared workspace agent | Business operations |
| Microsoft Copilot Studio | Employee or customer agent | Business operations |
| Salesforce Agentforce | CRM action | Business operations |
| Google Gemini Enterprise | Work across company apps and data | Business operations |
| Intercom Fin | Support conversation | Business operations |
| LangGraph | Stateful run | Frameworks and runtimes |
| CrewAI | Crew or workflow in code | Frameworks and runtimes |
| Google ADK | Agent with tools and workflows | Frameworks and runtimes |
| OpenAI Agents SDK | Agent, handoff and trace | Frameworks and runtimes |
| Claude Managed Agents | Session on managed infrastructure | Frameworks and runtimes |
| AWS Bedrock AgentCore | Execution infrastructure | Frameworks and runtimes |
When is another agent worth adding?
In Kim and colleagues’ study of 260 configurations across six benchmarks, the extremes ranged from +80.8% on decomposable financial reasoning to −70.0% on sequential planning, relative to the corresponding single-agent baseline. The figures come from different tasks and setups. Another study groups multi-agent failures into design issues, inter-agent misalignment and task verification. Scaling agent systems · Multi-agent failures.
Decomposable financial reasoning +80.8%
Sequential planning −70.0%
Two extremes from Kim and colleagues. Each comes from a different task and setup.A research dossier that spans several independent markets can be divided, with everyone using the same source format. If the answer to a customer depends on a single disputed contract, find and verify that source first. The right test is whether the final result takes less human time to accept.
What to check before you choose
Run the same small task on your shortlist. Write down which model and account are used, where the work runs, which files and tools the agent can access, and where it leaves its output. Change one input and ask for a revision. If the agent will take external actions, verify where approval takes effect and what happens if the call fails or gets retried.
Compare the cost per accepted result. Include model usage, platform fees, operator time, review and retries. Run logs can tell you that a process finished. Compare the deliverable with what the next person actually needs.
Keep building
Claude Code n8nVCG Ideas
Your next idea. And how to build it.
Join VCG Ideas for researched business ideas, recommended stacks and guides to building with AI. We’ll let you know when the first edition is published.
Join the list Free newsletter in Spanish. You can unsubscribe at any time.