alphadl/AgentHER
AgentHER: Hindsight Experience Replay for LLM Agents
What's novel
AgentHER: Hindsight Experience Replay for LLM Agents
Code Analysis
16 files read · 4 roundsA 4-stage pipeline that takes failed LLM agent trajectories, validates the failure, extracts actual achievements, reverse-engineers a new prompt the trajectory satisfies (with multi-judge verification), and packages the result into SFT/DPO/ShareGPT training formats.
Strengths
Clean separation of concerns with dual-mode (rule-based/LLM) operation at each stage, thoughtful multi-judge verification algorithm with temperature escalation and confidence merging, and robust structured-output parsing with multiple fallback strategies. The code is well-typed, well-documented, and the CLI is polished with rich output.
Weaknesses
No tests for the most complex stage (prompt_relabeler), and the rule-based failure detector's keyword matching is fairly naive (e.g., 'i believe' as a hallucination signal). The async client is defined but never used in the pipeline, suggesting incomplete integration.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
44
Contributors
1
Files
29
Active weeks
3
Repository
Language
Python
Stars
1
Forks
0
License
Apache-2.0