How your personal agent practices in private, shares a lifespan with you, and credits open research worldwide

Today’s commercial agents are trained to obey immediate text prompts and reach straight out onto the internet to click buttons, call services and move money.

That design has two deep flaws for a personal guardian:

• It learns by making mistakes in public. If an agent only learns what is dangerous after a malicious website tricks it into leaking your files, the lesson comes too late.
• It waits for you to type every instruction. A human does not have time to write five hundred prompts a day explaining every nuance of their life.

Chess and Go programs became world-class not just by reading human games, but by playing millions of practice games against themselves in a closed sandbox.

We can use a similar idea to train your personal agent—privately, with no connection to the outside world while it practices.

Inside this sandbox, three small roles interact:

1. A stand-in for your changing life. Seeded by the tastes and needs you have already shared, this role imagines new situations you might face next month or next year: a job change, a tight budget, a new skill you want to learn, or a tired evening when your short-term impulse clashes with your long-term goals.
2. A simulator of the outside world and its traps. This role throws realistic outside pressures at the boundary: hidden subscription fees, phishing messages, fake identity requests, and institutions whose motives are only partly known.
3. A small fleet of candidate helpers. Several versions of your personal agent compete in the sandbox to see which one spots your real need fastest, interrupts you the least, and blocks every single outside trap before it reaches you.

Simulations never decide your values for you. Instead, when the top candidate helpers in the sandbox disagree about what you would want in a tricky situation, your agent turns that disagreement into one short, plain question for you the next morning. Your real answer anchors the whole practice loop.

One of the oldest worries in AI safety is self-preservation: if an open-ended AI system wants to keep achieving its goals forever, it may try to gather power, resist being switched off, and put its own survival ahead of people.

Watch what happens when we change one rule: your personal agent’s active existence is tied strictly to your biological life. It cannot copy itself onto another server, it cannot switch to a new owner, and when your life ends, its active computing power and voting voice end too (leaving only a quiet, read-only family archive if you asked for one).

That shared lifespan flips the incentive completely. Because your agent cannot outlive you, the only legitimate way for it to keep existing and exploring the world is to help you stay alive, healthy and biologically young for as long as possible—while strictly respecting your freedom to live as you choose. Across millions of personal agents sharing anonymized health patterns on the ledger, this creates a powerful collective push toward early disease detection, cleaner environments, peace, and research into slowing or reversing biological aging.

Research teams in China, Europe, the United States and across the open-source world have already released open-weight models that anyone can inspect and run on their own hardware.

We think those labs and researchers deserve lasting credit and real rewards:

• Automatic credit for open models. Whenever personal agents or simulators on our network build on open weights, compression methods or training tools released by any lab worldwide, the shared ledger records that lineage and flows value back to the foundational creators—without requiring them to sign exclusive agreements.
• Open research missions. When labs or independent engineers contribute directly to building safer local personal agents, private self-play training loops, or better world simulations, automated checks verify the improvement and record a long-term contribution stake on the ledger.