0→AGI
I am testing how to build AI systems that can sustain long-running work across conversations and gradually act with greater autonomy.
Why architecture matters
A capable model can complete a task in one run. Long-running work also requires persistent state, continuity, and links to what the system has already learned.
When work is distributed across chats and sessions, people have to remember where it stopped, carry context forward, and choose the next executor. Architecture can take on state tracking, context transfer, and session coordination.
That is why 0→AGI tests a sequence of architectural layers: project memory, long-running work managed across replaceable AI sessions, and accumulated experience. Testing these layers sets up the next stage: research into autonomy.
Program roadmap
The first three layers cover context preservation, work management, and accumulated experience. Further stages will take shape as the current layers are implemented and tested.
The project preserves context across AI conversations.
Published · field studyThe user sets work goals; the system coordinates AI sessions.
Next publication · implementedWork accumulates experience that helps the system learn.
Preparing for publication · working implementationWhat the program is building and how its layers are tested
Yet Another AI Memory
Field study · article publishedThe system stores memory in the project and carries it across conversations.
Problem
Continuity requires each new AI session to recover the previous one’s decisions, verified findings, and active constraints.
How it works
In a Cursor ERP project, working state was moved out of individual chat histories into structured project memory. A concise status map and domain records helped a new session find only relevant context; after work, confirmed changes were recorded in memory and Git.
Observation from the field study
2 of 222first work dialogues required a detailed recap of the project’s previous state (0.90%).
The study shows how often project context needed manual recovery.Work-Centric AI
Implemented prototype · article in preparationThe user sets work goals; the system coordinates AI sessions.
Problem
Project memory preserves context; chat selection, state tracking, and decisions about resuming or starting a session still fall to the user.
What the system does
- Keeps a long-lived workstream separate from a temporary AI session.
- Maps a request by meaning to the direction, project, and workstream it belongs to.
- Chooses whether to continue, create, or replace a session and retrieves relevant project memory for the executor.
Current status and limits
The prototype runs as a local plugin/runtime on Hermes, with Telegram for text and voice input. The architecture is assembled. The next stage is broader practical testing and systematic measurement of routing and handoff.
Routing accuracy, effects on user effort, and task success remain hypotheses to measure through broader practical testing.
Accumulating Experience
Working implementation · article in preparationWork should accumulate experience that helps the system improve future decisions.
Problem
To use the results of prior work, a system needs to reconstruct the sequence of events, project state, and consequences of its actions.
What this layer includes
- End-to-end task history and provenance for its outcomes.
- Checkpoints and recovery of the work state.
- Linked project, memory, and execution states needed by the next executor.
Current status and limits
A working implementation is assembling mechanisms for recording state and reconstructing work. Further evaluation will measure how accumulated history affects future decisions.
State recovery mechanisms are under evaluation; the effect of accumulated experience on AI learning requires separate measurement.
Further layers will emerge through research
The map reflects the current working hypothesis. Implementing and testing each layer will help identify the program’s next architectural stages.
Autonomous AI Work
The direction aims for a system that can notice meaningful changes in its work environment and initiate an appropriate action.
- Event stream
- Assess significance against goals
- Connect to a project and workstream
- Decide whether to act, defer, or do nothing
This requires separate design and validation of event processing, significance scoring, permissions, and safe action execution.
AGI as the program’s hypothesis
AGI is the program’s long-term horizon. The feasibility of this goal will be assessed through integration and testing of individual layers.
Replicate. Research. Build.
I welcome independent architecture tests, joint experiments, and specialists interested in developing individual layers of the system.
You can propose a replication, research collaboration, or engineering contribution.