The Pushary Manifesto
Let your agents work. Get the decisions that need you. Pushary will be the standard decision layer for AI agents, across the tools you use today and tomorrow.
Pushary will be the standard decision layer for AI agents.
We are building for the agents people use today and the ones nobody has invented yet. Models will change. Frameworks will change. The need to decide who has authority will remain.
A decision layer lets an AI agent ask the right person, get an answer, and keep a record before the work continues. Rules handle decisions you have already delegated. New decisions come back to you.
Sometimes the decision is permission. Sometimes it is a choice, missing context, or a change of direction. An agent needs a dependable way to get all of them.
1. Intelligence does not grant authority
An agent can know exactly how to issue a refund. That does not tell it whether your business should issue this one.
It can write a convincing email. That does not give it permission to send it to your customers.
It can be right about the next step and still have no authority to take it.
Better models will solve harder problems. We want them to. But a prediction about what someone would approve is not their approval. Capability and permission are separate questions, and they stay separate as intelligence improves.
People should be able to delegate real authority: run these tests, work inside this project, handle these routine requests. That delegation needs boundaries they can understand, change, and withdraw.
Inside those boundaries, the agent should move. When the next action needs a new decision, it should ask.
More capable agents deserve clearer authority.
2. Stop making people babysit software
An agent that asks about every harmless step makes the person do a second job: supervising the machine.
After enough interruptions, approval becomes a reflex. The button is still there. The judgment is gone.
We want fewer questions, with more meaning. Let explicit rules handle the routine work. Bring people the decisions that need their judgment, with enough context to make the call: what will happen, what it affects, and why the agent needs an answer.
Then reach them where they are.
I built Pushary because my own agents kept waiting in windows I wasn't looking at. I was using Claude Code and Codex. The work could move quickly until an agent needed me. Then everything depended on whether I happened to check the right tab.
That is a bad way to run one agent. It gets worse with five.
Today, Pushary brings questions to the phone and the Mac notch, and carries the answer back. Permission policies let supported agent hooks handle routine actions under rules the user controls.
The measure of progress is useful work completed under the authority you gave it. Every unnecessary interruption gets in the way.
Ask when it matters. Make the answer count.
3. The decision belongs to the person affected
The person running an agent is not always the person entitled to decide.
A developer builds a purchasing assistant. The customer owns the budget. A support team runs a refund agent. The business sets the refund policy. An operations agent prepares a production change. The person responsible for that system decides whether it goes live.
Human-in-the-loop software has to reach beyond the developer's terminal.
That is why Pushary has a Partner API. Builders can connect their own users, send them a decision, and receive the answer. Their backend checks that answer before carrying out the action. They can use Pushary's delivery surfaces or collect the answer in their own signed-in application.
An agency should be able to put a prepared customer message in front of the business owner who can approve it. A product team should be able to reach its customer's designated reviewer when that person has closed the dashboard. Neither should have to build a mobile app just to get an answer.
A notification saying something happened is useful. A decision that arrives before it happens gives the person a say.
Sending the question is only the beginning. The answer must return to the right agent. A rejection must be respected. Silence is never a yes. There must be a record of the outcome.
We build for that complete loop.
Your agent asks. The right person decides.
4. Authority must survive a change of agent
People will change models. Teams will use several frameworks. Products will replace the agents behind their interfaces.
Every switch should not mean rebuilding how a person grants permission, receives a question, or reviews what happened.
The decision layer must work across those choices. It must serve the developer running coding agents and the company building agents for its customers. It must be useful on a laptop today and in workflows that run for days tomorrow.
Pushary already connects through agent integrations, MCP, and APIs. Builders can use framework adapters to bring human decisions into their products. The code that runs an action must check permission first. Supported hooks do that for the actions they cover. API users enforce the answer in their own code. Installing an optional MCP tool alone does not block other tools.
Those boundaries must be explicit. A standard earns trust through behavior people can inspect.
We are building one consistent way to ask, answer, and keep the record across agents and products. Builders should be able to add Pushary once and keep it as the rest of their stack changes.
The tools will change. Your authority should carry through.
We are building the standard
Asking the right person must become a basic capability of every agent that acts on someone's behalf. We intend to make Pushary the default way builders give their agents that capability.
For the individual, that means leaving the desk while the work continues within the boundaries they set. For the builder, it means adding human decisions without building the entire delivery and response system. For the customer, it means having a real say in what software does for them.
We will earn that place by making decisions easy to request, easy to understand, and dependable to act on. By reducing unnecessary interruptions. By making a refusal as real as an approval. By treating a person's time and authority as things worth protecting.
Pushary will be the standard decision layer for AI agents. We are building it now, for a future with far more agents doing far more consequential work.
Agents will do more of the work. People will own the decisions.
Connect your agent to Pushary, or build human decisions into your product.
For the product decisions behind this ambition, read why we built Partner APIs.