AI agents that get better by doing the work
We're building self-improving agents that learn from real tasks, not just training runs. Starting with coding because it's measurable. Engineering teams and researchers use our agents to compound their AI investment over time.
Most AI stops learning after training.
We think that's the wrong end state.
The current paradigm trains models on massive datasets, then freezes them. Users get a static snapshot. We're exploring a different path: agents that continue to improve through work, feedback, and iteration.
Why coding first?
Software work is measurable, auditable, and self-contained. Tests pass or fail. Code ships or breaks. It's the perfect sandbox to prove whether an agent can plan, act, learn, and genuinely improve—not just simulate improvement.
The compounding loop
Every task is a learning opportunity. Agent executes work, observes outcomes, refines behavior. The loop compounds over time. This is how capability should grow—through use, not just scale.
Work → Feedback → Improvement
This is the research engine. Not theoretical—running in production, closing real tickets, getting better at it.
HAL is the proof
HAL is our live agent—operating in real repositories, managing real workflows, demonstrating the thesis in practice. Built on Clawdbot. Learning every day.
Persistent memory
Remembers context across sessions. Learns preferences over time.
Tool orchestration
Spawns sub-agents, manages tmux sessions, coordinates multi-step work.
Real workflows
Integrated with Phorge, Git, messaging. Does actual work, not demos.
The roadmap
Starting personal. Scaling to teams. Then science labs. The architecture is designed to grow.
HAL Personal
Single-user AI assistant with full autonomy. The proving ground for the improvement loop.
HAL for Teams
Shared agents, shared memory. Coordination across humans and AI working together.
ScienceOS
Multi-agent research automation. From idea to published paper. The full loop at scale.