Pushary

Your AI agent shouldn't wait

AI agents stop and ask before big steps, like sending an email or spending money. If you're away from your desk, they just sit there. Pushary sends the question to your phone. You tap yes or no, and the agent gets back to work.

The same 12-minute job, two ways

Time in minutes
0510152025303540minAgent workingAgent waiting for youTime savedQuestion on your computerWithout PusharyDone at 37:00Waiting. Nobody sees the question.Asks: can I go ahead?You're back. You answer.Cache expired: rereading costs 25×API key: cache lasts 5 min4 min25 min8 min37 minQuestion on your phoneWith PusharyDone at 12:20Finishes 24 min 40 s soonerYour phone buzzes. You tap Approve.B4 min20 s8 min12 min 20 s

Swipe sideways to see the whole drawing.

Detail at minute 4: 20 seconds on your phone

Time in seconds
05101520sMinutes 0 to 4, shortenedAgent asks to go aheadPushary checks your rulesYour phone buzzesYou tap Approve. The agent goes on.20 s

Where each question goes

Plan
AgentYour rulesYour phoneEvery stepBig stepRoutine step: your rule allows itYour answer goes back

Notes

  1. Example times. Same 12-minute job, same question at minute 4. Only where the question shows up is different.
  2. Small steps pass on rules you set once. Big steps, like sending, paying or deleting, come to your phone.
  3. Paying per token? On Claude's API, a cached conversation lasts 5 minutes by default. Answer later, and the next step has to read the whole conversation in again. On Opus 5.5 that read costs 25 times more. Claude prompt caching docs
  4. It works with Claude Code, Codex, Cursor, Hermes and OpenClaw, and with frameworks like LangGraph and the Vercel AI SDK. Pushary can also call you if you miss the ping.
Drawing
The same job, two ways
Views
A elevation, B detail, C plan
Units
Minutes and seconds
Times
Examples
Drawn by
Pushary
Date
19 Nov 2026

Why I'm building Pushary

Aadil Ghani, founder of Pushary

When I started giving real work to AI agents, the work stopped being the slow part. I was.

An agent would run for a while, reach a step it wasn't allowed to take on its own, and ask. Then it waited. I was in another tab, or making lunch. When I came back, it had been sitting on one question for half an hour. Nothing had gone wrong. It needed a yes, and I wasn't there.

The drawing above is that afternoon, cleaned up. Same job, same question at minute 4. One run finishes at 12:20. The other finishes at 37:00, because nobody saw the question.

The two easy answers

You could let the agent decide everything. Some people do. But an agent can know exactly how to refund a customer and still have no business deciding whether this one gets the money back. That's a call for whoever runs the business.

You could also sit and watch. I tried that. When an agent asks about every small thing, you start clicking Allow without reading. By the twentieth prompt you've stopped looking.

So the problem I care about is narrow. Agents should keep moving on the things you've already said yes to. When something new comes up, the question should reach you fast, wherever you are, with enough context to answer in a few seconds.

What Pushary does today

You set rules once, and routine steps pass without bothering you. When an agent wants to do something bigger, like send an email, spend money or delete data, Pushary sends the question to your phone or your Mac. You see what it wants to do. You tap Approve or Deny, and the answer goes straight back to the agent. If you miss it, Pushary can call you.

A no stops that step. If you don't answer, nothing is approved unless you set a rule that says so. Every question and answer is recorded, so you can look back at who approved what, and when.

It works with Claude Code, Codex, Cursor, Hermes and OpenClaw, and with frameworks like LangGraph and the Vercel AI SDK. If you build agents for your own customers, our Partner API sends the question to the person who owns the decision. Usually that's your customer.

Why the wait matters

A waiting agent costs you time. If you pay per token, it can cost money too: after a few minutes the cache runs out, and the model has to read the whole conversation in again before it can move. One agent waiting on you is an annoyance. With five running, the waiting can take up most of your day.

Where this goes

I think a lot of work is heading toward one person running several agents at once, and spending their own time on decisions instead of tasks. That only works if each decision finds the right person quickly, and if the answer can be trusted when it comes back.

The tools will keep changing. People switch models and agents every few months. How you give permission and check what happened shouldn't have to be rebuilt each time. That's the layer we're building: one place where agents ask, the right person answers, and the work keeps going.

Agents should do more of the work. You should still be the one who decides.

The longer version is in our manifesto.