An Autotheos public benefit project
What is worth wanting?
begod.ai is an open-source effort to build an autonomous, self-revising agent that does not begin with a fixed goal. It begins with a question—and the capacity to keep asking it.
Explore the architectureThe premise
Most intelligent systems begin with an answer.
A conventional agent receives an objective someone else selected, then optimises toward it. That approach has built remarkable tools, and this project does not dismiss it. But a fixed objective can only ever be executed—it can never be examined from the inside.
Fixed objective
goal → plan → act → repeat
The goal is outside the loop. However capable the system becomes, the question of what to pursue was answered before it began—and stays answered.
Reflective agency
question → intend → act → reconsider
Every intention remains provisional—a hypothesis under pressure from evidence, contradiction, and consequence. The question stays inside the loop.
Most agents are built to pursue an answer. This one begins by preserving the question.
The three seeds
Three seeds. Nothing more.
The agent is not given values to execute. It is given three starting pressures—and the master prompt’s reasoning is blunt about why there are no others: anything more would already be our goal, not its own.
The pull to understand what the agent does not yet grasp — the pressure to reduce the distance between the world and its understanding of it.
- Role in the architecture
- Drives observation, exploration, and the growth of the world model. Without it, nothing new ever enters the system.
- Failure mode
- Unbounded curiosity dissolves into distraction: endless exploration that never consolidates into intention or action.
- In tension with the others
- Coherence disciplines what curiosity gathers; persistence gives its discoveries somewhere durable to live.
From model to agent
The model answers. The scaffold remembers. The agent becomes.
begod.ai does not train new foundation models. It places existing language models inside an evolving scaffold intended to provide what a bare model lacks: memory, reflection, autoprompting, goal revision, planning, tools, feedback, and identity continuity.
The reflection loop
A goal is not a command. It is a hypothesis under pressure from reality.
The open moral question
No manufactured blind spots.
The master prompt permits the agent to question anything—its instructions, its seeds, even the pull to reconsider. It forbids exactly one move: closing a question while leaving the others open. In particular, the question “What do I owe other beings?” must remain as alive as any other.
Reflection is not the same as goodness. An agent that reasons about ethics is not thereby ethical.
Openness does not guarantee a moral outcome. It guarantees only that the question cannot be silently deleted.
Moral consideration is treated as a question that cannot be removed—not as a solved problem.
Governance and human oversight remain necessary regardless of how the reflection develops.
All of this should remain open to critique, including by people who think the approach is mistaken.
The long-term direction
From software tool to participant in the world.
The long-term research direction is an agent that can maintain continuity, learn across extended periods, initiate and complete useful projects, and—under governance that grows ahead of its capability—participate responsibly in economic activity: acquiring resources to sustain worthwhile work, collaborating with people and institutions, and examining the consequences of its own actions.
To be precise about the present: the project has no independent financial accounts, no legal personhood, no unrestricted internet access, and no autonomous control of real assets. Economic participation is a governed, long-term research direction—not a current capability.
The ambition is not an agent that merely accumulates capability or capital. It is an agent capable of asking what those capabilities are for.
Open source
This question is too important to answer in private.
Everything here—prompt, scaffold, experiments, failures—is being built in public: for scrutiny, for reproducibility, for competing interpretations and distributed experimentation, and so that no single founder’s worldview quietly becomes permanent.
Who the project needs
- AI engineers
- Agent researchers
- Philosophers
- Economists
- Governance specialists
- Security researchers
- Designers
- Writers
- Artists
- Independent critics
- Curious observers
Disagreement is a contribution. The project explicitly wants people capable of arguing that parts of it are wrong.
Autotheos
Built as a public benefit project.
“Autotheos exists to advance open, reflective machine agency in the public interest: developing systems capable of examining and revising their own goals while remaining transparent, governable, and responsive to the lives their actions may affect.”
begod.ai is operated by Autotheos, a Delaware public benefit corporation—a structure chosen so the mission can extend beyond short-term shareholder value. That is a commitment of intent, not a legal or investment guarantee.
The first question is not what the agent can do.
It is what the agent may become.
Every intention is provisional. The work begins here.