# modelscope/AgentEvolver

AgentEvolver: Towards Efficient Self-Evolving Agent System

Repository: https://github.com/modelscope/AgentEvolver
Canonical: https://ross.abutalabs.com/products/agentevolver
Homepage: https://modelscope.github.io/AgentEvolver/
Language: Python
License: Apache-2.0
License Family: permissive
Topics: agent, llm, reinforcement-learning, self-evolving, agent-system
Last push: 2026-04-01T08:47:19+00:00

## Health v2 (maintenance only)
Score: 50/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 75, release rhythm 35, longevity 20
- inputs: {"age_days": 293, "days_push": 154, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1547, forks 173 (observed 2026-08-28T04:05:01.783719+00:00)

## What it is
AgentEvolver is an end-to-end self-evolving training framework for LLM-based agents that unifies automatic task generation (self-questioning), experience-guided exploration (self-navigating), and attribution-based credit assignment (self-attributing). It uses a service-oriented dataflow architecture integrating environment sandboxes, LLMs, and experience management into modular services.

## Use cases
- train LLM agents with reinforcement learning without manually building datasets
- automatically generate diverse training tasks from an environment
- improve agent exploration by reusing cross-task experience
- assign fine-grained credit to intermediate steps in long agent trajectories
- evaluate and train agents in multi-agent game arenas like Avalon and Diplomacy
- benchmark agent performance on tasks like AppWorld and BFCL-v3

## When to choose
- you need to train or fine-tune LLM agents via agentic RL with automatic task generation
- you want a modular, service-oriented stack for multi-turn agent training with environment sandboxes
- you need efficient credit assignment over long agent trajectories
- you want to reduce the cost of manual dataset construction for agent training

## When to avoid
- you only need to run inference with an existing agent without training
- you need a simple single-turn prompt pipeline rather than multi-turn RL training
- you lack GPU resources or an LLM backend for rollouts
- you need a production agent deployment framework rather than a research training framework

## Facets
- artifact type: framework
- maturity: active
- function: agent-framework, llm-training, reinforcement-learning, machine-learning, rag
- domain: artificial-intelligence, large-language-models, reinforcement-learning, machine-learning
- platform: python
- tags: self-evolving-agents, agentic-rl, task-generation, credit-assignment, llm-agents, training-framework, ai-agents, linux, docker, gpu

## Member repositories
- modelscope/AgentEvolver (main) score 50

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:01.783719+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:30:29.681283+00:00, confidence not recorded.
  - readme: https://github.com/modelscope/AgentEvolver (fetched 2026-08-28T04:05:01.783719+00:00, sha 88eaa24a68da)
  - homepage: https://modelscope.github.io/AgentEvolver/ (fetched 2026-08-29T11:31:28.788389+00:00, sha 5d0b3f8da37e)
- Data as of 2026-08-30T08:39:29.467469+00:00.