System-4
A multi-turn, steerable, long-horizon autonomous system for recursive agent work. System-4 is a research preview.
Explore the research ↗Small team working on the hardest problems in AI
Mythein builds System-4.
A swarm-scaling superintelligence OS that evolves its own harness.
Mythein Desktop
Hand it a hard problem: post-train Qwen3.6-35B-A3B for Finance Agent v2.
Fleet stage. The fleet in six crews: who is running, done and queued, and what each one was asked.
A sample session recorded in Mythein Desktop's test mode. The task comes from Vals AI's RSI Index; its baseline (35.38%) and reference (59.42%) are Vals AI's published figures, and every other number is illustrative.
64 agents, one desktop.
Mythein Desktop puts System-4’s Swarm Computer on your machine. Hand it a hard problem and a lead agent plans the work, splits it across crews and runs a fleet of up to 64 agents, while you watch every task, wait and reply and steer the run as it goes.
What the fleet makes stays with you. Dashboards, tables, reports, protein structures and code changes open beside the chat and are kept in Artifacts; automations rerun saved prompts on a schedule.
Continuous Skills. When a run works out a method worth repeating, its lead offers to keep it as a skill: a short SKILL.md with the steps and when to use them, saved to your library or to the project. Every agent in the fleet can follow it in later runs, so the next run starts from what the last one learned. Nothing is kept until you say so.
SuperChips
A massively parallel RL infrastructure for chip design that cooperates with System-4.
Visit SuperChips ↗Recursion happens in the harness.
System-4 works on the machinery around an agent: its instructions, tools, skills, scripts and orchestration. It proposes task-specific changes, probes them in execution, and composes a candidate harness from the resulting traces.
Adopted harnesses and lessons are retained as versioned experience for future tasks. The system’s working method is something it can revise and carry forward—not a fixed wrapper around a model.
System-4’s objective is stronger generalization across tasks and faster convergence as its harness evolves.
Harness refinement · next task
Edit the system that does the work.
Change task-specific instructions, tools, skills, scripts, and orchestration—not just the output.
An operating system for recursive agents.
System-4 is a domain-general framework for recursively improving agent harnesses. It brings task-level refinement, durable workflows, evidence checks and versioned experience into one system.
Superhuman intelligence is the ambition. We make no claim that System-4 has achieved it.