Prime Agent
Open-source self-improving agent harness that spawns its own sub-agents. Described in an August 2026 paper as a self-improving RLM (recursive language model) harness.
Features
- Spawns its own agents — recursive delegation rather than a fixed agent graph
- Self-improving loop: the harness revises its own strategies across runs
- Model-agnostic — runs with Ollama, LM Studio, DeepSeek V4 Flash, Claude, Codex
- Free to run locally; “free forever” setups documented
Notes
Prime Agent belongs to the late-2026 wave of open agent harnesses — alongside
DeepSeek Harness and NanoClaw — that compete on being free, local and
model-agnostic rather than on a proprietary model. It was explicitly benchmarked as beating
Hermes Agent in August 2026.
Young tool, first documented August 2026. Capability claims are untested over any meaningful
duration.