Releases·Tool Release
Nvidia Released SoL-Pi, a Pi Extension Packaging Four Auto-Research-Discovered Agent Harness Efficiency Mechanisms
Nvidia Labs released SoL-Pi, a standalone extension for the Pi agent that packages four efficiency mechanisms discovered via scaled auto-research loops. It cuts repeated model turns, context replay, oversized observations, and long-log reading while preserving task-critical work and evidence. Practitioners can install it on top of an unmodified Pi release to reduce spend without reducing useful agent work.
1 engagement1 sourceThu, Sep 10, 2026, 08:17 PM
AI Overview
- SoL-Pi is a standalone Nvidia extension that installs on unmodified Pi via public APIs, reducing repeated turns, context replay, oversized observations, and long-log reading
- Its four opt-in mechanisms: Action Fusion (edit plus validation in one call), ObservationPack handles, Evidence-Preserving Reducer log receipts, and Online Context Compact of completed plan steps
- Design constraints: no Pi patches, explicit opt-in with mechanisms disabled by default, locally preserved evidence, and Pi retains control of model and runtime choices
Related
Developments·Paper
SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness
An arXiv paper introduces SoL-Pi, an agent harness discovered by scaling recursive auto-research loops across many environments. Four mechanisms survived selection, covering action execution, context compaction, observation handling, and delegated reading. On EdgeBench, SoL-Pi matches Pi's performance while cutting recorded token traffic roughly in half and API cost by about a third, offering practitioners measurable hourly savings on unattended coding-agent workloads.
An arXiv paper introduces SoL-Pi, an agent harness discovered by scaling recursive auto-research loops across many environments. Four mechanisms survived selection, covering action execution, context compaction, observation handling, and delegated reading. On EdgeBench, SoL-Pi matches Pi's performance while cutting recorded token traffic roughly in half and API cost by about a third, offering practitioners measurable hourly savings on unattended coding-agent workloads.
0 engagement1 sourcearxivThu, Sep 17, 2026, 02:58 PM
Releases·Tool Releasefeatured
Prime Agent open-source RLM harness scores 95.5% on ARC-AGI-3
Prime Agent, an open-source coding and research agent built on pi, claims a 95.5% score on ARC-AGI-3, surpassing the human-expert baseline. The harness is designed for token efficiency and long-running autonomous tasks, with improvements across models compared to proprietary harnesses.
Prime Agent, an open-source coding and research agent built on pi, claims a 95.5% score on ARC-AGI-3, surpassing the human-expert baseline. The harness is designed for token efficiency and long-running autonomous tasks, with improvements across models compared to proprietary harnesses.
368 engagement1 sourceredditWed, Aug 5, 2026, 11:30 PM
Products·launched
Infinite context agent session without compaction
Pi-CWL is an agent framework that runs a single session through over 80 million tokens without needing context compaction, avoiding information loss and hallucination. It solves the finite context window problem for long-running agent tasks, targeting developers building complex multi-step agent workflows.
Pi-CWL is an agent framework that runs a single session through over 80 million tokens without needing context compaction, avoiding information loss and hallucination. It solves the finite context window problem for long-running agent tasks, targeting developers building complex multi-step agent workflows.
2 engagement1 sourcehackernewsTue, Jul 14, 2026, 04:26 PM
Releases·Tool Release
Show HN: Felan – an open source coding agent focused on efficiency
Developers released Felan, a free open-source coding agent built as an opinionated version of Pi. It targets lower coding costs through model routing, context management, and compact tool output, supporting many providers via both API usage and subscriptions. For practitioners, it offers a self-hosted way to cut token spend while keeping agents task-effective.
Developers released Felan, a free open-source coding agent built as an opinionated version of Pi. It targets lower coding costs through model routing, context management, and compact tool output, supporting many providers via both API usage and subscriptions. For practitioners, it offers a self-hosted way to cut token spend while keeping agents task-effective.
2 engagement1 sourcehackernewsTue, Sep 8, 2026, 06:31 PM
News·Company News
GitHub cuts Copilot costs with output compression, prompt trimming, and notification batching
A GitHub engineering post outlines how the team makes AI coding more cost efficient without sacrificing task quality, promoting GitHub Copilot CLI for agentic workflows in the terminal. It argues efficiency should be measured by completing tasks with the right amount of context, not by minimizing token counts per interaction.
A GitHub engineering post outlines how the team makes AI coding more cost efficient without sacrificing task quality, promoting GitHub Copilot CLI for agentic workflows in the terminal. It argues efficiency should be measured by completing tasks with the right amount of context, not by minimizing token counts per interaction.
0 engagement1 sourcerssWed, Sep 2, 2026, 06:00 PM