SoL Pi

SoL Pi is an AI agent system scales recursive auto-research to reduce agent cost and token use while preserving task quality.

SoL-Pi scales recursive auto-research to reduce agent cost, tokens, and turns while preserving useful work, evidence, and task quality. It treats system improvement as an open-ended recursive self-improvement problem, searching across different environments for changes that can transfer to unseen settings. The auto-research process generates mechanism ideas in batches, sends each selected idea into an independent loop, and carries forward candidates that survive validation. It uses parallel auto-research loops, a Ralph Loop for iterative implementation, map-reduce analysis for exploration trajectories, and two isolated splits for validation. The project builds 535 executable training environments in two families, including tasks mined from GitHub issue-PR pairs and synthetic tasks defined by executable verifiers.

SoL Pi at a glance

Best for: Software engineer, Researcher
Runs on: Web, open source
Pricing: Free
Released: September 2026
Category: AI research tool
Website: nvlabs.github.io

Key features

Scales recursive auto-research to reduce agent cost and token use
Uses parallel auto-research loops for independent mechanism lineages
Employs a Ralph Loop for iterative implementation until explicit exit conditions pass
Performs map-reduce analysis across multiple exploration trajectories
Validates candidates using two isolated splits for training and held-out test sets
Builds 535 executable training environments from GitHub issues and synthetic verifiers

Use it for

Reducing agent token use and cost during unsupervised coding tasks
Scaling recursive self-improvement for AI software systems
Evaluating agent performance on held-out test sets using executable verifiers

Companions

Tools that articles about SoL Pi recommend using with it, each with its job.

Codex — Built to work with Codex.
Context7 — Provides up-to-date documentation for AI agents and code editors.
Firecrawl — Provides direct web data access for AI agents with open source flexibility.
Superpowers — Plan, test and execute workflows across harnesses.
Impeccable — Audits and improves UI design, typography and accessibility.
RTK — Compresses command output before it reaches the context.

Alternatives to SoL Pi

OpenResearch — Run parallel research workflows locally or on remote compute.
Composio — Equips AI agents with tools, managed user auth, and sandboxes across 1,500+ apps
pstack — Improves AI coding agent outcomes with rigorous workflows.
Tura — Reduces token usage and LLM turns by over 80 percent while improving performance.
AI Scientist v3 — Automates end-to-end scientific research from task descriptions to complete papers.
Auto Deep Researcher — Automates experiment operations to reduce manual babysitting of training loops.
AutoResearch — It provides a stateful, reproducible workflow for automated AI research.
Rsi Eval — It supplies the evaluation boundary so methods can use existing benchmarks without reimplementing execution.
GEAR — Organizes recursive self-improvement through evolutionary search and reproducible evaluation.
Beeuni — Agents retain context and improve skills over time through a controlled review process.
deepsense.ai — A framework to evaluate agentic systems before production usage exists.
AI Research Skills — 98 expert skills for training, evaluating and writing up ML research

All SoL Pi alternatives