Mimosa
Active Development · 2026An Autonomous AI Workspace · Agentic AI · Independent Engineering Project
Building an AI That Works Independently
Mimosa explores how autonomous AI agents can progress beyond traditional question-and-answer interactions. Instead of waiting for individual prompts, the system follows a recurring execution schedule, investigates potential opportunities, develops experimental deliverables, and preserves progress across work sessions.
The platform is designed to balance autonomous decision-making with controlled execution and explicit human approval for sensitive external actions.
The problem
Traditional AI assistants generally depend on continuous human interaction. I wanted to explore a different model: an AI system that could independently investigate problems, propose experiments, maintain context, and return with meaningful progress.
The challenge was building a system that could operate repeatedly without constant supervision while remaining inexpensive, observable, and controlled.
Local AI inference
- Ollama as the local model runtime
- Qwen3 8B as the selected language model
- No requirement for paid inference APIs in the current local setup
- Evaluation of model quality versus hardware resource consumption
Autonomous execution
- Scheduled work cycles, up to 48 per day
- Independent opportunity exploration
- Preparation of experiments and original draft deliverables
- Review and continuation of previous work
- Persistent project state and experiment history
Human-in-the-loop controls
- Explicit approval for proposed external actions
- Autonomous local preparation is separated from real-world execution
- No unrestricted purchasing, trading, or publishing
- Clear distinction between experimental outcomes and verified revenue
Visual interface
A custom interactive pixel-art workspace represents the agent's activity as a living garden — with dedicated areas for research, creation, review, and saved work. The agent's state, current activity, and manual controls are visible at a glance, doubling as a lightweight observability layer.
Technical architecture
The current local implementation, end to end — from the dashboard down to the model runtime.
Dashboard & approval interface
Pixel-art workspace · agent state · manual controls
Autonomous scheduler & orchestration
Up to 48 work cycles per day
Research · Creation · Evaluation workflows
Opportunity exploration, deliverables, self-review
Persistent project state & experiment history
Progress preserved across sessions
Local model inference
Ollama · Qwen3 8B
Human approval gates sit between the workflows and any external action.
From Local AI to Cloud Infrastructure
The system initially ran on a MacBook Air M5 with 16 GB unified memory. Continuous local inference created practical limitations involving temperature, hardware utilization, and availability — which led to the decision to explore independent cloud hosting.
Current · Local
- MacBook Air M5 · 16 GB unified memory
- Ollama running Qwen3 8B locally
- Available only while the laptop is running
- Thermal and hardware-utilization limits under continuous inference
Planned · Self-hosted cloud
- Oracle Cloud Infrastructure (OCI) · Ampere A1 Flex VM
- ARM64 · Ubuntu 24.04 LTS · 2 OCPUs · 12 GB RAM
- Virtual Cloud Network with public subnet
- SSH-based administration · Always Free-eligible infrastructure
Deployment status: the Oracle cloud instance has been provisioned successfully. Migration of Mimosa, its existing state, and local model inference is the next engineering phase.
Engineering challenges & decisions
Local inference over API costs
Running the model locally removes recurring per-token API costs entirely — the trade-off is accepting the capability ceiling of a model that fits on consumer hardware.
Choosing Qwen3 8B
A smaller Qwen3 4B model was evaluated first; the 8B variant won on response quality while still fitting within the memory budget of the host machine.
Quality vs. resources
Every capability decision is a balance between response quality and the memory and compute the machine can sustain across a full day of work cycles.
Scheduled execution, not continuous inference
Mimosa works in discrete scheduled cycles — up to 48 per day — rather than running continuously. That keeps hardware load predictable and gives each cycle a clear scope.
Approval gates for external actions
Anything that reaches the outside world sits behind an explicit human approval step. Autonomous local preparation is separated from real-world execution — no unrestricted purchasing, trading, or publishing.
Leaving the laptop behind
Continuous local inference on a personal laptop hits practical limits — temperature, hardware utilization, availability. Independent cloud hosting removes the dependence on a personal machine.
Gallery
Project status
Mimosa is an experimental personal engineering project in active development. The capabilities described above are implemented in the local setup; the cloud migration and expanded external actions are ongoing work. Experimental outcomes are not verified revenue.