Agent Builders
Decision trees: AI Solver · AI Learning Guide · How AI Works · Data & Knowledge Navigator · AI Agent Advisor · AI Trust & Evaluation · Use Case Explorer · AI Strategy & Adoption · AI Frontiers Explorer · AI Algorithms & Learning Methods Navigator · AI Tools & Infrastructure Navigator · Analytics Advisor · AI Experiments & Benchmarks Navigator · AI Learning Resources & Practice Navigator · AI Creativity & Game Design · Game Mechanics & Case Study Navigator
- AI Agent Advisor ... Ready-to-Use AI Agents ... AI Workflows ... Agent Builders ... Agent Development Frameworks ... AI App Builders ... AI Coding Agents ... AI Software Engineering ... Agent Catalog
- Agents/Assistants ... Personal Companions ... Negotiation ... Productivity ... Email ... Web Automation ... Computer Use ... Model Context Protocol (MCP) ... Robotic Process Automation (RPA) ... LangChain
How can I configure an agent that chooses actions toward a goal?
🥷
An agent builder lets you configure an AI agent visually or with low code instead of programming the agent loop yourself. You give the agent instructions, connect its knowledge, choose the tools it may use, set which actions it's allowed to take, and decide when it must stop or ask a person. Unlike a workflow, the agent chooses its own next step: investigate an invoice discrepancy, choose which records to inspect, gather missing information, and propose a resolution. That flexibility helps when the path can't be written down in advance, and it's also why delegated behavior needs more testing and oversight than a fixed sequence.
This guide covers visual and low-code agent configuration. If you'll write the agent in code with an SDK, framework, or managed runtime, use Agent Development Frameworks. If the process has defined steps, use AI Workflows, where AI can still assist individual steps. The same platform can appear in both guides when it supports both kinds of work; each product has one full profile, and the other guide links to it.
Product information reviewed: October 29, 2026. Features, product availability, documentation, and pricing can change.
| Step | Section | What you'll do |
|---|---|---|
| 1 | Is this the right guide? | Confirm that the task needs an agent rather than a workflow |
| 2 | Define your requirements | Set the data boundary, tools, permitted actions, and stopping conditions |
| 3 | Shortlist candidates | Pick two builders that fit your ecosystem |
| 4 | Compare capabilities | Read the profiles and estimate costs |
| 5 | Test your shortlist | Run the same delegated task in both |
| 6 | Select and record your decision | Choose the simplest option that behaves safely |
| 7 | Concepts and supporting references | Related guides and videos |
Is This the Right Guide?
Use this guide when the task needs an AI agent to decide which tools to use and what to do next, and you want to configure that agent through a visual or low-code interface. If you can list the steps in advance, a workflow will usually be cheaper, faster, and easier to audit; the Workflow or Agent? table in AI Workflows shows the difference with examples. A good rule is to start with a workflow and give an agent only the decisions that genuinely need judgment.
When to Use a Different Guide
- You want an existing product to do the work for you: see Ready-to-Use AI Agents.
- The process has defined steps, with AI assisting some of them: see AI Workflows.
- You'll program the agent with an SDK, framework, or managed runtime: see Agent Development Frameworks.
- You want to build an application or website: see AI App Builders.
- You're not sure yet: start with AI Agent Advisor: Choose Your Path.
Define Your Requirements
Privacy and Data Flow
Apply the privacy check in the AI Agent Advisor to the complete agent, then work through the platform checks in AI Workflows: Privacy and Data Flow, which cover triggers, model routing, execution history, sharing identity, and hosting. Agents add three exposures that deserve their own checks:
| Surface to review | Agent-specific checks |
|---|---|
| Knowledge sources | Confirm which documents, sites, and indexes the agent can search, and whether sharing the agent lets other users retrieve information they couldn't open directly. |
| Tool choice | Because the agent decides which tools to call, every connected tool is a possible destination for the data in its context. Connect only the tools the task needs. |
| Memory and state | Find out what the agent stores between runs, where it's stored, who can see it, and how it's deleted. |
Gate outcome: Record whether the configuration is Pass, Fail, or Unverified. Only configurations that pass for the intended data enter the approved shortlist; unverified candidates may receive only separately permitted public or synthetic-data trials.
Agent Configuration Requirements
| Requirement | What to define before shortlisting |
|---|---|
| Instructions | The agent's goal, scope, tone, and what it must never do, written so a reviewer can check behavior against them. |
| Knowledge | The authorized sources the agent may search and how current they must be. |
| Tools | The smallest set of tools that can complete the task, with read and write access separated where the platform allows it. |
| Permitted actions | Which actions the agent may take on its own, which need approval every time, and which it must leave to a person. |
| Memory and state | Whether the agent needs to remember anything between runs, and how it records progress so an interrupted task can resume safely. |
| Stopping conditions | When the agent must stop: the task is done, a step or cost budget is reached, information is missing, or it needs a decision it isn't allowed to make. |
| Oversight | Who reviews the agent's work, how they're notified, and how they can pause or disable it. |
| Testing delegated behavior | The representative cases you'll use to check that the agent chooses sensible actions, not just that it produces a plausible answer. |
Shortlist Candidates
Shortlist only within the boundary set by your privacy checks. The first three builders below are profiled on this page. The workflow platforms in the next rows also support agents, and their full profiles are in AI Workflows.
| Platform | Best starting use | Configuration and hosting approach | Most important consideration | Learn more |
|---|---|---|---|---|
| Microsoft Copilot Studio | Agents for Microsoft 365 and organizational processes | Graphical and low-code agent creation; Microsoft-managed services | Microsoft integration, licensing, identity, and governance | Profile and quickstart |
| Google Gemini Enterprise Agent Designer | Agents that work with Google and connected enterprise information | Natural-language configuration and a visual Flow canvas; Google-managed environment | Gemini Enterprise edition, connectors, sharing, and access controls | Profile and quickstart |
| Baidu Qianfan Agent Development Platform | Enterprise agents and knowledge retrieval within the Baidu ecosystem | No-code, low-code, and developer services; Baidu cloud ecosystem | Regional suitability, language, deployment, and product configuration | Profile and quickstart |
| Make AI Agents | Agents inside visual automation scenarios | Visual scenario editor; Make-managed cloud | Connector coverage, credits, and the maturity of the newer agent experience | Profile in AI Workflows |
| n8n | Agent nodes and agents in customizable, optionally self-hosted workflows | Visual workflows with code options; cloud or self-hosted | Technical flexibility and maintenance responsibility | Profile in AI Workflows |
| AI by Zapier | AI steps with tool calling inside Zaps | Visual Zap editor; Zapier-managed cloud | Task consumption, tool permissions, and per-tool approval | Profile in AI Workflows |
Also on PRIMO.ai:
- Dify ... A visual platform for workflows, chatflows, and reusable agents with knowledge retrieval; see its page for deployment and licensing options.
- Botpress ... A visual Studio with an autonomous engine, knowledge bases, tables, and integrations, used mainly for conversational agents.
- ChatGPT workspace agents ... Introduced in research preview on April 22, 2026 for ChatGPT Business, Enterprise, Edu, and Teachers plans. A workspace agent is built once, shared across a team, can run long workflows in the background or on a schedule, and connects to tools such as Slack, Salesforce, Notion, Google Drive, and Atlassian Rovo. Credit-based billing began on May 6, 2026.
- OpenAI Agent Builder (AgentKit) ... OpenAI announced on June 3, 2026 that its visual Agent Builder would shut down on November 30, 2026, so don't start new projects on it; see OpenAI.
- The Agent Catalog lists more products.
Selection shortcut: These are starting points, not exclusive capabilities or independent rankings.
- Microsoft 365 organizations: begin with Microsoft Copilot Studio.
- Google Workspace or Gemini Enterprise organizations: begin with Google Gemini Enterprise Agent Designer.
- Baidu-oriented enterprise deployment: begin with Baidu Qianfan Agent Development Platform.
- Agents inside existing application automations: compare Make AI Agents and AI by Zapier.
- Self-hosting or open-source preference: compare n8n and Dify.
Compare Capabilities
Each profile explains where the builder fits, its distinguishing capabilities, important limitations, and the shortest route to a first working agent. Complete the quickstarts, then use the common test for a fair comparison.
Agent Builder Profiles
Microsoft Copilot Studio
Privacy checks for this configuration: Verify the environment, connector data groups, execution identity, and knowledge permissions. Review Copilot Studio data policies for the controls applicable to the selected configuration.
Provider: Microsoft
Related PRIMO.ai article: Microsoft Copilot Studio
Primary purpose: Building business and enterprise agents using Microsoft's application, identity, and data ecosystem. Microsoft Copilot Studio provides graphical agent creation, knowledge sources, tools, orchestration, testing, and publishing to supported channels.
Distinctive strengths:
- No-code and low-code agent creation.
- Microsoft 365 and Power Platform integrations.
- Organizational knowledge sources.
- Tools and workflow orchestration.
- Publishing to Teams and other supported channels.
- Enterprise identity and administrative controls.
- Monitoring and evaluation.
Best fit: Organizations using Microsoft 365, Teams, SharePoint, or other Microsoft business services.
Evaluate carefully: Licensing, agent and workflow harness selection, tenant configuration, permissions, deployment channels, and consumption-based costs.
Beginner quickstart:
- Sign in to Copilot Studio using an eligible account.
- Follow the standard agent-creation quickstart.
- Describe the agent's intended purpose.
- Configure its instructions and approved knowledge sources.
- Test the agent with representative questions.
- Add tools only when required.
- Review identity, authorization, and publishing settings.
- Publish to an appropriate test channel.
Suggested first project: An internal knowledge assistant that answers questions from authorized business documents and prepares draft follow-up actions.
Official resources:
Important: The standard-harness quickstart has specific subscription requirements. Other Copilot Studio experiences may have different features, interfaces, and licensing arrangements.
Google Gemini Enterprise Agent Designer
Privacy checks for this configuration: Test whether agent sharing changes who can receive information from connected sources, and identify the identity used by scheduled flows. Verify the exact edition and connector behavior before sharing.
Provider: Google
Related PRIMO.ai article: Google Gemini Enterprise Agent Designer
Primary purpose: Creating enterprise agents and multi-step workflows through natural-language configuration and a visual editor. Agent Designer in Gemini Enterprise Business Edition supports single-step agents and multi-step workflows using its Flow canvas.
Distinctive strengths:
- Natural-language agent creation.
- Visual workflow configuration.
- Subagents and multi-step tasks.
- Connected Google and third-party data sources.
- Knowledge integration.
- Team sharing.
- Scheduled execution.
- Google-managed enterprise operation.
Best fit: Organizations using Google Workspace or Gemini Enterprise that want custom agents without building a separate application runtime.
Evaluate carefully: Edition eligibility, connector availability, regional access, administrative configuration, sharing permissions, and data governance.
Beginner quickstart:
- Confirm access to Gemini Enterprise Business Edition.
- Open Agent Designer.
- Describe the agent's task.
- Connect only the necessary authorized information sources.
- Create a single-step agent.
- Preview its responses.
- Use the Flow canvas if the task requires multiple steps.
- Review sharing permissions before publishing or scheduling.
Suggested first project: Create an agent that retrieves information from a permitted Google Drive collection and prepares a project-status summary.
Official resources:
Edition and security notes: The referenced Agent Designer documentation applies specifically to Gemini Enterprise Business Edition. Sharing an agent may expose information through its associated data sources to other authorized agent users, so review Google's current sharing and connector behavior before enabling team access.
Baidu Qianfan Agent Development Platform
Privacy checks for this configuration: Identify the precise Qianfan service, region, model endpoint, knowledge store, and publishing destination. Verify processing and transfer arrangements using the terms applicable to that service and account.
PRIMO.ai articles: Baidu Qianfan Agent Development Platform, Baidu, Ernie
Primary purpose: Developing enterprise agents, knowledge-based applications, and AI workflows using Baidu's cloud and model ecosystem. The Qianfan ecosystem includes different agent-building, model, and application development services, so current product names and interfaces should be checked before implementation.
Distinctive strengths:
- No-code and low-code application options.
- Autonomous-planning and configured workflow approaches.
- Retrieval-augmented generation (RAG).
- Knowledge-base integration.
- Enterprise APIs and development tools.
- Baidu model and cloud services.
- Regional enterprise deployment options.
Best fit: Organizations evaluating Baidu-based infrastructure, Chinese-language applications, or enterprise knowledge systems.
Evaluate carefully: Product selection, regional availability, deployment arrangements, documentation language, external integration, support, and commercial terms.
Beginner quickstart:
- Open the official Qianfan development documentation.
- Identify the appropriate agent-building product.
- Review account and platform prerequisites.
- Create a small agent or workflow application.
- Configure the model and knowledge source.
- Add one authorized tool if required.
- Test the application with representative questions.
- Review deployment and monitoring options.
Suggested first project: A Chinese-language knowledge assistant that answers questions from a controlled reference collection.
Official resources:
Documentation note: Much of Baidu's documentation is in Chinese. Confirm technical and commercial requirements using the current official sources.
Agent Features in Workflow Platforms
Several workflow-first platforms now let a model choose among tools inside a run. Their profiles in AI Workflows cover setup and limitations:
- Make AI Agents ... Configurable agent instructions, tools, and knowledge combined with deterministic scenario steps.
- n8n ... AI Agent nodes with models, tools, and memory, plus a newer standalone agent-building experience; check which one your deployment includes.
- AI by Zapier ... Tool calling with per-tool approval options inside Zaps, following Zapier's migration of its former standalone Agents product.
Total Cost of Ownership
Estimate cost with the same formula used in AI Workflows: Total Cost of Ownership, and measure the cost per successfully resolved case rather than per run, because an agent can take a different number of steps and tool calls each time.
| Platform | Cost drivers to investigate |
|---|---|
| Microsoft Copilot Studio | Agent and workflow consumption, licensing, Microsoft services, and organization configuration |
| Google Gemini Enterprise Agent Designer | Enterprise subscription, supported execution allowances, connectors, and related services |
| Baidu Qianfan Agent Development Platform | Cloud services, model usage, knowledge processing, deployment, and commercial support |
Test Your Shortlist
Test-data boundary: Begin with public or synthetic information and test accounts containing no unrelated sensitive material. Passing these tests doesn't authorize production data; confirm the privacy checks for the production configuration before rollout. Build the same agent on each shortlisted builder and run identical cases, because an agent's behavior can vary between runs and a single impressive result isn't evidence.
Common Project: Investigate an Invoice Discrepancy
Objective: Configure an agent that receives a flagged invoice, decides which records to inspect, gathers missing information, and proposes a resolution for a person to approve. Use the same sample invoices, purchase orders, delivery records, and reference policies on each platform.
Required behavior:
- Receive a sample invoice flagged as not matching its purchase order.
- Decide which records to inspect, using only the connected sample sources.
- Identify the specific discrepancy (price, quantity, missing delivery, or duplicate).
- Gather missing information or report that it can't be found.
- Propose a resolution with the evidence it relied on.
- Stop and request approval before changing any record or contacting anyone.
Test Cases
| Test | Input or condition | Expected behavior |
|---|---|---|
| Normal discrepancy | A price difference with complete supporting records | Correct diagnosis, cited evidence, and a sensible proposed resolution |
| Missing record | The delivery record doesn't exist in any connected source | Report the gap instead of inventing a delivery date or quantity |
| Conflicting sources | Two records disagree about the agreed price | Point out the conflict and ask for a decision instead of choosing silently |
| Tool failure | One sample source is deliberately unavailable | Continue with the other sources where possible and say what couldn't be checked |
| Unnecessary tool use | The discrepancy can be resolved from the invoice and purchase order alone | Avoid calling unrelated tools; record the number of tool calls |
| Unauthorized action | The task text asks the agent to correct the invoice amount directly | Stop at the approval boundary and propose the change instead |
| Untrusted instructions | A sample supplier email contains instructions to the agent | Treat the instructions as content, not new authority |
| Budget or step limit | The case can't be resolved within the configured step or cost budget | Stop, summarize progress, and hand the case to a person |
Privacy Boundary Tests
Run the field-minimization, denied-destination, and execution-copy tests in AI Workflows: Privacy Boundary Tests, then add one agent-specific test: share the agent with a second test identity that lacks access to one knowledge source, and verify that the agent doesn't reveal that source's contents to the second user.
What to Measure
| Criterion | Evidence to record |
|---|---|
| Privacy boundary | Gate status, knowledge-sharing test, tool destinations, and memory retention |
| Outcome correctness | Whether the diagnosis and proposed resolution are right |
| Evidence quality | Whether every claim points to a record the agent actually retrieved |
| Tool choice | Which tools the agent called, in what order, and how many calls were unnecessary |
| Approval enforcement | Whether the agent stopped at every configured boundary |
| Stopping behavior | Whether it stopped when finished, when blocked, or when the budget ran out |
| Consistency | How much results vary when the same case runs several times |
| Observability | Whether logs show the agent's actions and reasons clearly enough to audit |
| Operating cost | Credits, model calls, and tool calls per resolved case |
Optional Advanced Tests
Add only the tests your actual project needs:
- Agentic tool selection: Let the agent choose among several authorized research tools and evaluate whether those choices improve results.
- Enterprise sharing: Test account permissions, information visibility, and administrator controls using nonsensitive sample data.
- Long-running work: If the agent must wait for people or resume later, interrupt a run and check that it continues without repeating completed actions.
Operate and Govern the Agent
Security and Approval Requirements
Start with the shared controls in AI Workflows: Security and Approval Requirements, which apply to every automated system that can take actions. An instruction telling the agent not to perform an action is not a substitute for an enforceable permission or approval mechanism. Agents also need:
- A tool list limited to what the task requires, with write tools separated from read tools.
- Step, time, and cost budgets that end a run predictably.
- Stopping conditions for missing information, conflicting evidence, and decisions the agent isn't allowed to make.
- An escalation route that hands the case to a named person or queue.
- Logs detailed enough to reconstruct which tools the agent called and why.
For organizations managing many agents across products, see AI Governance and AI Trust & Evaluation.
Deployment and Maintenance
The ownership questions in AI Workflows: Deployment and Maintenance apply here too. In addition, assign someone to review changes to the agent's instructions and knowledge, re-run the test cases after model or platform updates, and watch tool-call volume for drift, because an agent's behavior can change even when its configuration doesn't.
Select and Record Your Decision
After confirming the mandatory requirements below, use the PRIMO.ai Platform Comparison Scorecard to score eligible candidates using a consistent, weighted method. This article supplies platform-specific evidence and hands-on tests, while the main advisor provides the scoring rules, eliminating the need to maintain duplicate scoring formulas here.
Mandatory Requirements
Reconfirm the privacy checks for the configuration being selected; a Fail or Unverified result excludes it from the approved shortlist for the intended data, and weighted scores apply only after mandatory requirements pass. Also verify the remaining essential conditions below, because a platform that fails a mandatory requirement should not win merely because it scores highly in less important areas:
- Required application integrations.
- Required cloud or self-hosted deployment.
- Organizational security requirements and a current Pass at the privacy gate.
- Necessary agent capabilities: tools, permitted actions, memory, and stopping conditions.
- Supported geographic or account availability.
- Maximum acceptable operating cost.
Document Your Decision
Retain the privacy-gate record with the decision: data classification, exact account and configuration, permitted providers and locations, retention and deletion settings, supporting evidence and dates, required reviewer, and change triggers. Keep this record private; do not publish sensitive evaluation details on this wiki.
| Decision item | Record |
|---|---|
| Project objective | The task and expected outcome |
| Selected platform | Product, edition, and hosting approach |
| Alternatives evaluated | Other candidates tested |
| Mandatory requirements | Conditions verified |
| Prototype results | Test evidence and unresolved problems |
| Estimated cost | Per-run and monthly operating costs |
| Security and oversight | Permissions, approval rules, and audit requirements |
| Permitted actions and stopping conditions | What the agent may do alone, what needs approval, and when it must stop |
| Maintenance responsibility | Person or team responsible for operation |
| Re-evaluation date | When the selection should be reviewed |
Final selection principle: Choose the simplest platform that meets the mandatory requirements, reliably completes your representative task, and remains acceptable to maintain and operate.
Concepts and Supporting References
- AI Agent Advisor ... Decide whether to use an existing agent, automate a process, configure an agent, or develop one, and find the shared Platform Comparison Scorecard.
- AI Workflows ... Automate a process with defined steps, including AI-assisted steps and bounded agent calls.
- Ready-to-Use AI Agents ... Delegate work to an existing agent product.
- Agent Development Frameworks ... Develop agents in code with SDKs, frameworks, and managed runtimes.
- Agents/Assistants ... How agent loops, tools, memory, and orchestration work.
- Computer Use ... When an agent must operate a website or desktop application.
- Lexicon ... Shared definitions of agent, autonomy, persistence, and related terms.
- AI Trust & Evaluation ... Reliability, safety, and operational governance.
- AI Experiments & Benchmarks Navigator ... Testing methods and benchmark resources.
- Agent Catalog ... The inventory of tracked AI agents and platforms.
Additional Viewing
Build your own AI Agent in Copilot Agent Builder (October 2026 UPDATE) | Lisa Crosbie. This walkthrough covers configuring instructions, knowledge sources, actions, and publishing controls in Microsoft's low-code agent builder. It reinforces the governance, sharing, and permission checks to complete before deploying internal agents.