LN: Xie et al. (2023) — OpenAgents: An Open Platform for Language Agents in the Wild

Bibliographic Reference

Citation: Xie, T., Zhou, F., Cheng, Z., et al. (2023). OpenAgents: An open platform for language agents in the wild. arXiv:2310.10634. https://arxiv.org/abs/2310.10634


Pass 1 — Bird’s Eye View (5 Cs)

CAssessment
CategorySystem design + deployment paper
ContextAddresses the gap between academic agent benchmarks and real-world deployable agent platforms
CorrectnessDeployed publicly with user evaluation; provides usage statistics from real users
Contributions(1) Three specialized agents: Data Agent (pandas/SQL), Plugins Agent (200+ tools), Web Agent (browser automation); (2) Open-source implementation; (3) Natural language interface for non-technical users
ClarityGood. Each agent is described independently with concrete examples.

Relevance: ⭐⭐⭐⭐

The Data Agent architecture (NL → Python/SQL → execution) directly maps to PUMA’s pipeline for processing Jira datasets.


Pass 2 — Content

Three Agent Types

AgentPrimary ToolsUse Case
Data AgentPython, pandas, matplotlib, sklearnData analysis, visualization, ML pipelines
Plugins Agent200+ plugins (weather, maps, finance…)Real-world task automation
Web AgentBrowser automation, web scrapingInformation gathering, form filling

Data Agent Architecture (most relevant to PUMA)

User Query (NL)
    ↓
Code Generator LLM → Python/pandas code
    ↓
Safe Execution Sandbox (subprocess isolation)
    ↓
Output + Error Handling (retry on exception)
    ↓
Response to User

The Data Agent generates Python/pandas code, executes it in a sandboxed subprocess, captures stdout/stderr, and feeds errors back to the LLM for self-correction — essentially implementing a code-based Reflexion loop.

Key Design Patterns

  • Tool use through code generation: Instead of tool APIs, agents write code — more flexible but requires safe execution environment
  • Error recovery: LLM receives the Python traceback and rewrites the code
  • Conversation memory: Previous steps are included as context for multi-turn analysis

PUMA Integration

  • Dataset processing: The Data Agent architecture (NL → pandas → results) is the pattern for PUMA’s data preprocessing pipeline for TAWOS and Jira SR
  • Stage 2 automation: Automate data quality checks, stratified sampling, and feature extraction using a Data Agent pattern
  • Open-source foundation: OpenAgents is MIT-licensed; PUMA could fork the Data Agent for interactive Jira dataset exploration

MOCs