Trending Model:#1Inklingthinkingmachines⬇16kTrending Model:#2Unlimited-OCRbaidu⬇2237kTrending Model:#3Ternary-Bonsai-27B-ggufprism-ml⬇432kTrending Model:#4Laguna-S-2.1poolside⬇3kTrending Model:#5Bonsai-27B-ggufprism-ml⬇1405kTrending Model:#6Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUFDavidAU⬇63kTrending Model:#7GLM-5.2zai-org⬇545kTrending Model:#8Nanbeige4.2-3BNanbeige⬇0kTrending Model:#9Solar-Open2-250Bupstage⬇0kTrending Model:#10Qwen3.6-35B-A3B-Uncensored-HauhauCS-AggressiveHauhauCS⬇1998kTrending Model:#1Inklingthinkingmachines⬇16kTrending Model:#2Unlimited-OCRbaidu⬇2237kTrending Model:#3Ternary-Bonsai-27B-ggufprism-ml⬇432kTrending Model:#4Laguna-S-2.1poolside⬇3kTrending Model:#5Bonsai-27B-ggufprism-ml⬇1405kTrending Model:#6Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUFDavidAU⬇63kTrending Model:#7GLM-5.2zai-org⬇545kTrending Model:#8Nanbeige4.2-3BNanbeige⬇0kTrending Model:#9Solar-Open2-250Bupstage⬇0kTrending Model:#10Qwen3.6-35B-A3B-Uncensored-HauhauCS-AggressiveHauhauCS⬇1998k

Autoswarm Forges Self-Improving Assistants from Local Language Models

By vramkickedin | | 2 min read
A self-organizing swarm of luminous mechanical nodes forming a neural mesh.

Autoswarm is a new open-source tool that turns local language models into self-improving assistants. It works as a drop-in API proxy that logs your chats and distills useful lessons into a skillbook, then injects those skills into future prompts. The project also includes a benchmark mode where a multi-agent pipeline rearranges its own structure to solve tasks better.

Arteemg built autoswarm to solve two problems: static LLMs that never learn from past interactions, and hand-crafted agent pipelines that require constant human tuning. The online proxy mode allows any OpenAI-compatible client to benefit from automatic skill learning and pruning, while the benchmark harness lets a meta-agent iterate on pipeline topology to maximize task completion. The tool targets privacy-conscious users who want their local models to improve without sending data to the cloud.

Self-improving proxy and pipeline optimization

Key Features
  • Drop-in proxy for Ollama, LM Studio, vLLM.
  • Auto-generates and prunes a skillbook from chats.
  • Multi-agent pipelines that edit their own topology.
  • Per-stage evaluation pinpoints weak pipeline steps.
  • CLI commands for doctor, reflect, and skills.
  • OpenAI-compatible endpoint on localhost:8080.
  • Benchmark mode hill-climbs on passed tasks.
  • Works entirely with your existing local LLMs.

The proxy is perfect for tinkerers who want their chatbots to gradually get better at answering domain-specific questions. Small teams can run it in front of their internal LLMs to automatically capture company knowledge without manual prompt updates. Advanced users will appreciate the benchmark mode, which provides a playground for evolving custom agent workflows and comparing structural changes.

How the meta-agent edits pipelines

The benchmark mode relies on a coding agent that reads program_pipeline.md and modifies benchmark/pipeline_spec.yaml, the sole edit surface. After each run, a per-stage judge scores every handoff, and the meta-agent uses these scores to decide which stages to adjust, add, or remove. All structural changes are logged in results.tsv, so users can trace how the pipeline topology evolved over experiments.

"A self-improving OpenAI-compatible proxy for local LLMs, plus a multi-agent pipeline harness that self-optimizes its own topology." — Source: GitHub

Project Links