Ways to keep a human in the loop

Five approaches, ranked

The ranking assumes your agent runs unattended and its steps can matter. If you sit at the machine the whole time, the order changes. We make Notifex, so read the first card with that in mind.

1. Option 1

   A push-based question inbox (Notifex)

   Best when you are away from the keyboard while the agent runs.

   The agent calls an MCP tool and asks. Your iPhone gets a push notification, you answer, and the agent resumes. Questions can be single-select, multi-select or free text, and up to 10 can arrive in one push. Agents can attach images, video and documents. Every thread stays in one inbox.

   Strengths

   • You answer from anywhere, not only at your desk
   • Structured answers instead of free-form replies to parse
   • Team projects: the first answer wins

   Trade-offs

   • Needs the iPhone app and one MCP entry per agent
   • iPhone only today
   • Another app to keep notifications on for
2. Option 2

   Chat or email messages

   Best when you already live in one chat or mail client and answers can be slow.

   A script or workflow posts the question to a chat channel or sends an email, and a reply is read back. It reuses a tool you already open. The cost is the glue: someone builds the sender, the reader and the matching of replies to questions.

   Strengths

   • No new app on your phone
   • Familiar threads and notification settings

   Trade-offs

   • You build and maintain the reply matching
   • Questions mix with unrelated messages
   • Free-text replies need parsing
3. Option 3

   A custom approval web UI

   Best when you need a tailored screen, roles or an audit trail you control.

   You build a small web app where the agent files a question and a person approves or rejects it. You control every field and rule. You also own hosting, sign-in, notifications and the work of getting someone to open it in time.

   Strengths

   • Full control of layout, rules and data
   • Can fit unusual review steps

   Trade-offs

   • You build, host and maintain it
   • Still needs a way to tell the person a question is waiting
   • Time to first working version is the longest of the five
4. Option 4

   Terminal prompts

   Best when you are sitting at the machine and watching the run.

   The agent stops and asks in the terminal where it runs. Nothing to set up, and the context is right there. The limit is presence: if you step away, the run waits on a prompt you cannot see.

   Strengths

   • Zero setup
   • Full context on screen

   Trade-offs

   • Blocks until you return to that terminal
   • No notification when a question appears
   • Hard to share with a teammate
5. Option 5

   Auto-approve with logs

   Best for low-risk, easy-to-undo steps you can review afterward.

   The agent never stops. Every decision is written to a log and you read it later. It is the fastest option and removes the human from the moment of decision. If a choice cannot be undone, a log tells you what happened after it is too late.

   Strengths

   • Fastest: nothing ever waits
   • A record of what the agent did

   Trade-offs

   • No chance to say no before the action
   • Mistakes are found after the fact
   • Wrong for steps that are hard to reverse

How to choose

Start with the cost of a wrong guess, then match the channel to where you will be.

• Reach: can you answer when you are away?

  If the agent runs while you are out, a channel that pushes to your phone matters more than one you must open.

• Reversibility: how bad is a wrong guess?

  The harder a step is to undo, the more it needs a person before it runs, not a log after.

• Setup and upkeep: who builds and maintains it?

  A custom screen or a chat bridge is code you own. A terminal prompt costs nothing. Weigh that against how often you will use it.

• Answer shape: do you need structured replies?

  Single-select and multi-select answers are easier for an agent to act on than a free-text reply it has to interpret.

• Team: who else should be able to answer?

  If more than one person can unblock the agent, decide who answers first and who sees the result.