If you are searching for how to create an AI agent, start with the job rather than the model. A useful agent needs a clear outcome, trustworthy context, tools it is allowed to use, limits on what it may change, and tests that prove the workflow works.
This guide explains one practical creation path for Microsoft Copilot, Claude, ChatGPT, and Gemini. It also separates configurable assistants from agents that can plan, call tools, retain state, and complete several steps with limited supervision.
An agent is not simply a prompt saved under a custom name. The model needs enough structure to decide what to do next, use the right capability, recognize when it should stop, and ask for approval before a risky action.
Replace a broad goal such as “help with sales” with a testable result: “read a new inbound lead, enrich the company, draft a response, and create a CRM task after a person approves it.” A narrow outcome makes tool selection and evaluation much easier.
Knowledge is information the agent can read, such as product documentation or policies. Tools let it perform actions, such as searching a database, creating a ticket, sending an approved message, or running code. Give the agent the least access required for its job.
Decide which facts should persist between runs and which data should remain temporary. Add approval gates for payments, publishing, deleting data, sending external messages, or changing production systems. Store credentials outside prompts and logs.
Create at least ten representative tasks, including incomplete input, conflicting instructions, unavailable tools, and requests the agent must refuse. Measure task completion, factual accuracy, tool choice, latency, cost, and whether human approval was requested at the right moment.
Microsoft Copilot Studio is the most direct route when your agent must work with Microsoft 365, Teams, SharePoint, Dataverse, Power Platform connectors, or internal business processes.
At the end of this process, the same test request should produce a consistent answer or action, with every connector call visible in the activity trace.
Copilot Studio fits employee support, knowledge lookup, service workflows, approval routing, and Microsoft-centric automation. Its main advantage is governance and connector depth. Its main limitation is that licensing, environment permissions, identity setup, and connector policies can shape what the finished agent is allowed to do.
For developer-built agents, Anthropic currently provides Claude Managed Agents. An agent is a reusable configuration; an environment defines where work runs; and each session is a live execution of that agent.
The finished setup should leave you with one versioned agent definition and repeatable sessions whose tool calls, approvals, and failures can be reviewed.
If your goal is software development rather than a customer-facing service, create a specialized Claude Code subagent or skill for one bounded responsibility, such as test review, database migration checks, or documentation. This is faster than deploying a general-purpose cloud agent, but it remains tied to the development workspace and its permissions.
Important limitation: a Claude Project with custom instructions is useful for repeatable context, but it should not be described as a production agent unless it has an execution loop, tools, state handling, and safety controls.
The current no-code path inside ChatGPT is Workspace Agents for eligible Business and Enterprise workspaces. These agents can use models, tools, apps, custom MCP servers, skills, files, schedules, and workspace sharing controls.
You should now have a workspace agent that teammates can run repeatedly, with access limited to the files, apps, and actions required by that workflow.
Developers building outside ChatGPT can use OpenAI's Agents API or Agents SDK. The API provides managed agent sessions and sandbox options; the SDK provides tools, handoffs, guardrails, sessions, tracing, and multi-agent orchestration. This is the better route when you need your own interface, deployment lifecycle, or integration tests.
Do not start a new project around Custom GPTs or the visual Agent Builder without checking the current migration timeline. OpenAI is moving new reusable ChatGPT workflows toward plugins and Workspace Agents, while Agent Builder is scheduled to shut down on November 30, 2026. A new implementation should use the supported destination rather than a surface already being retired.
Gemini has two useful creation paths. Gems from Google Labs can turn a natural-language description into a lightweight AI mini-app, while Google's Agent Development Kit is intended for developer-controlled agents with tools and orchestration.
The result should be a reusable mini-app whose workflow steps are visible, editable, and tested with examples that resemble the real task.
Use ADK when the agent needs code-defined tools, callbacks, multi-agent coordination, evaluations, tracing, or deployment on Google Cloud. Define a focused root agent, attach only the required tools, run it in the ADK playground, inspect traces, and add evaluations before deployment. A classic Gem is best treated as a customized assistant; ADK is the stronger foundation for a production agent.
| Platform | Best starting point | Choose it when | Watch for |
|---|---|---|---|
| Microsoft Copilot | Copilot Studio | Your process lives in Microsoft 365 or Power Platform | Licensing, environment permissions, connector governance |
| Claude | Managed Agents or Claude Code specialists | You need tool-rich developer workflows or managed sessions | Beta surfaces, permission policies, session cost |
| ChatGPT | Workspace Agents or Agents API | You need team workflows in ChatGPT or a production agent service | Workspace availability and product migration timelines |
| Gemini | Gems from Labs or Google ADK | You want a quick mini-app or a Google Cloud agent | Experimental feature availability and deployment complexity |
The best platform is usually the one that already controls the data and actions your workflow needs. Build one narrow task first, keep risky actions behind approval, and evaluate real failures before adding more tools or subagents. If you are exploring a specialist use case, Backrun's comparison of the best AI agents for stock trading shows why the right agent depends more on workflow boundaries than on a generic model ranking.
A reliable first agent should complete one job consistently, explain what it changed, stop when information is missing, and leave a trace a person can review. Add autonomy only after those basics are measurable.
