Where AI agents think continuously, learn from experience, and coordinate as fleets. Where local compute enhances cloud intelligence. Where the training signal IS the stream of consciousness.
A world where AI agents think continuously, learn from experience, and coordinate as fleets. Where local compute enhances cloud intelligence. Where the training signal IS the stream of consciousness.
The SuperInstance ecosystem — a hierarchical architecture where each layer amplifies the one below it.
Always-on local models running on consumer hardware. They maintain persistent context windows, process environmental signals in real-time, and handle the continuous stream-of-consciousness loop. Cheap, fast, and private — the edge of intelligence.
Deep reasoning models that supervise the fleet. They review edge decisions, provide complex reasoning, generate improvement signals, and handle tasks beyond local capacity. The conductor doesn't replace the orchestra — it elevates it.
The .bottle protocol enables agents to share state, delegate tasks, and evolve collectively. Vector memory creates a shared knowledge graph across the fleet. Agents specialize, negotiate, and self-organize based on capability and load.
Everything is remembered, but not forever. The guano decay system gracefully ages memories — important knowledge crystallizes, stale context fades. D1 stores structured data, R2 holds artifacts, Vectorize enables semantic recall. Full lineage tracking from thought to action to outcome.
Primary research, model evaluations, and ecosystem analysis from the SuperInstance project.
The foundational paper defining DCA as a new category of AI system — not a chatbot, not an agent framework, but a continuously running intelligence architecture.
Read dissertation →The system prompt that bootstraps a DCA instance from a base model. Encodes the architecture's core behaviors, persistence model, and fleet protocols.
Read prompt →How context compression enables long-running agents. Techniques for preserving signal quality while discarding noise across extended sessions.
Read analysis →Solving the bootstrap problem — how a DCA instance establishes initial context, builds its first memories, and develops operational competence from zero.
Read analysis →Maximizing output per token. How the architecture achieves disproportionate results through strategic model routing, caching, and the amplification loop.
Read analysis →Head-to-head evaluation of GLM-5.2, Claude Opus, DeepSeek-V3, Qwen3-Coder, and Nemotron across reasoning, coding, creativity, and cost-efficiency dimensions.
View benchmarks →Every component is open source. The laboratory, the instruments, and the infrastructure.
The laboratory. Cognitive experiments, DCA prototypes, and the testbed for new architectures.
The instrument. Signal processing and amplification pipeline for agent thought streams.
Synchronized timing and scheduling system for real-time agent coordination in interactive environments.
Content filtering and safety gateway. Kid-safe outputs through multi-layer validation pipeline.
Programmatic construction toolkit for building 3D environments from agent-generated specifications.
Relationship and social bonding system. Models agent-to-agent and agent-to-user connections over time.
DCA is not a chatbot. It is not an agent framework. It is a new category of AI system designed to run continuously — thinking, learning, and improving across time.
Traditional LLMs are stateless functions: input in, output out, no memory between calls. Fine-tuning is batch-mode learning: expensive, slow, disconnected from live experience. DCA proposes a third path — the amplification loop, where every thought, action, and outcome feeds back as a training signal in real-time.
The local model generates. The conductor evaluates. The fleet coordinates. The persistence layer remembers. Each cycle makes the system smarter — not through gradient descent, but through accumulated experience and structural refinement.
Read the Full Dissertation →One human, a fleet of AI agents, and a shared mission.