Build AI Agents You Still Approve

The AI agent builder is live on GoWithAgentic.ai. You describe the job in plain language, pick a model and a few tools, and run a test. The agent works out the steps. Every outbound action — an email, a reply, a task — lands in an Approvals inbox. A person on your team has to approve it before anything happens. Nothing is sent automatically.

This post is a practical walkthrough of what you can do today, how a test run looks, and what is still coming later. For the short version, start on the AI agent builder page or create a free trial.

What you can do today

After you sign in, open Agents. Give the agent a name and write the job the way you would brief a colleague. A typical brief for a first reply looks like this: when a new inquiry arrives, look up the contact if we already know them, draft a first reply that answers what they asked, offer a time to talk, and queue it for a person to approve. Never send the email yourself.

Then pick a model and the tools the agent is allowed to use. Five tools are live:

• Look up a contact — find a sample or CRM contact by name or email. It does not leave the workspace.
• Summarize — condense intake notes or a thread into a short summary. It does not send anything.
• Draft email — prepare an email. It waits in Approvals.
• Draft reply — prepare a reply to an inquiry. It also waits in Approvals.
• Create task — prepare a follow-up task. It is not created in a live system. That wait is the same as the email drafts.

Those five tools are the whole catalog right now. The agent chooses which ones to call and in what order. You do not draw a flowchart for each run. If a tool is switched off, the agent cannot use it.

Run a test and read every step

The builder has a Test run panel next to the brief. Paste a sample situation — for example, a new inquiry from Sarah about pricing for a 3-person clinic who wants a call next week — and click Run test.

The trace shows each step the agent took: a model turn, a contact lookup, a draft. Outbound tools are labeled as queued for approval. You can read the input the tool received and the payload it produced. If the agent drafted a reply, that draft is already sitting in Approvals. It has not gone to anyone's inbox.

This is the useful part. You are not trusting a black box. You are reading the same steps a teammate would have taken, then deciding whether the draft is good enough to keep. If the brief is vague, the draft will be vague. Tighten the job description and run the same sample again.

Test runs are manual. The agent does not wake up on a schedule, and it does not watch Slack. Those triggers are listed as coming soon below.

Three starter agents

You do not have to start from a blank brief. The builder ships three starters that match work most service businesses already do by hand.

New-lead reply drafter. When a new inquiry arrives, look up the contact if we already know them, then draft a first reply that answers what they asked and offers a time to talk. Queue the reply for a person to approve. Never send the email yourself.

Appointment reminder drafter. Given an upcoming visit (name, time, location), look up the contact and draft a reminder that asks them to confirm or reschedule. Queue the message for staff approval. Skip people who have already confirmed if the input says so.

Intake summarizer. Read intake notes or a new-inquiry form. Summarize facts, missing documents, and risks. Look up the contact if named. Draft a task for the next human step (conflict check, consult, or document request) and queue that task for approval. Do not contact the client.

Use a starter as written, or edit the brief. The job description is the control surface. If you do not want it to offer times, say so. If the reminder should never mention a fee, say so. The agent will still pick its own steps, but it stays inside the tools you enabled.

A new-lead agent pairs well with the BANT lead score when you want a transparent qualifier before a human spends time on the draft, and with the follow-up cadence planner when you already know which days a sequence should land. Those are separate, browser-based helpers. They do not send mail either.

If you are still deciding whether a process should be an agent or a fixed workflow, use the agent vs workflow scorer. Messy email is a good agent job. A reminder that must fire at the same time every night is a workflow job — and schedules for agents are not live yet.

Approvals: a person still decides

Approvals is a dedicated inbox in the logged-in app. Each item shows the action type (draft reply, draft email, create task), a short summary, and the payload. Filter to pending items, then Approve or Reject.

Approving marks the item approved. It does not send the email, create the calendar event, or write to a live CRM. Sending after approval is not enabled yet. Rejecting records the decision and still sends nothing.

That is intentional. The product rule does not change: nothing leaves the workspace until a person says yes, and today's yes is a record, not a send. If you want a checklist of which actions should always wait for a human, the HITL gate picker is a free way to name those gates before you automate them.

The trial uses sample contacts. Do not paste a real customer's private details into a test run unless you are ready for that text to sit in your workspace.

Coming soon

These are on the roadmap. They are not live. We are not attaching dates.

• Schedules — run an agent on a cadence instead of only from a manual test.
• Slack triggers — start a run from a Slack message.
• Call an agent from a workflow or API — a visual workflow (or an HTTP call) hands a job to an agent and waits on Approvals.
• Real sending after approval — email, calendar, and CRM writes that execute only once an item is approved.

Until those ship, treat the builder as a place to design the job, watch the steps, and practice the approval habit. The drafts are real enough to edit against. They are not going out.

Try it this afternoon

Start a free trial — 14 days, no credit card, sample data included. After you are in, open Agents, pick a starter, paste a sample inquiry, and run a test. Then open Approvals and decide the draft.

Read the full product page at gowithagentic.ai/ai-agents. Questions go to [hello@nexamartech.com](mailto:hello@nexamartech.com).