Agentic RL: Credit Assignment and CLI Agents

Measured conclusion: Deep-tested. L2 outcome: tools_ok.

Filter agent RL methods by supervision, critic and task setting; retrieve source links and BibTeX.

Add to your agent

Copy this into your MCP client config:

{
  "mcpServers": {
    "Agentic RL: Credit Assignment and CLI Agents": {
      "url": "https://hoyant-su-agentic-rl.hf.space/gradio_api/mcp/",
      "transport": "streamable-http"
    }
  }
}

Measured tools (8)

• Agentic_RL_list_sources

  List original papers and retrieval coverage. Discover source-linked comparisons of credit assignment, agent memory, selective observation and terminal benchmarks, with JSON, CSV and BibTeX links.

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• Agentic_RL_search_evidence

  Search original papers on agentic reinforcement learning, credit assignment and CLI agents. Use English keywords (AND), OR and quoted phrases. Return relevant passages, source citations, equations and table cells.

  Parameters:

  • query (required) — string
  • limit — integer

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• Agentic_RL_fetch_evidence

  Fetch a complete original evidence block by the evidence_id returned from search_evidence, including section anchor, version, equations, table cells, links, and attribution.

  Parameters:

  • evidence_id (required) — string

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• Agentic_RL_dataset_overview

  Inspect ShellOps and ShellOps-Pro task counts, train/test splits, task types, published schemas, source files, license and citation.

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• Agentic_RL_search_tasks

  Find real ShellOps CLI benchmark tasks by case-insensitive literal substring in the complete instruction, task ID or published task type. Empty query lists all tasks. Select partition 'all', 'shellops' or 'shellops_pro'; select published split 'all', 'train_src', 'train' or 'test'. Results are ordered by partition then task ID, with explicit pagination and no relevance scoring. The train subset is not double-counted.

  Parameters:

  • query (required) — string
  • limit — integer
  • split — string
  • offset — integer
  • partition — string

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• Agentic_RL_get_task

  Inspect one published ShellOps or ShellOps-Pro task by its exact task_id and partition ('shellops' or 'shellops_pro'). Returns the complete instruction, actual reward specification, published reference answer/command, file-entry metadata, pinned parquet rows and workspace asset links. File content is available at the source links. No shell execution or solution verification is performed.

  Parameters:

  • task_id (required) — string
  • partition (required) — string

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• Agentic_RL_list_method_facets

  List exact filter values for agent RL credit granularity, supervision, value critics and evaluation settings. Each value reports its source-supported method count.

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

• Agentic_RL_filter_methods

  Filter agent RL credit-assignment methods by research conditions and return original section evidence and BibTeX. Discover accepted values with list_method_facets. Filters combine with AND; empty strings leave a facet unrestricted. Unknown critic status never matches no. Results use publication order without a relevance or quality ranking.

  Parameters:

  • credit_granularity — string
  • evaluation_setting — string
  • learned_value_critic — string
  • required_supervision — string

  Hints: read-only not declared · destructive not declared · idempotent not declared · open-world not declared

Publisher website

Measured by MCPYP probing. Untested fields are omitted, never estimated. Listing is always free; MCPYP never proxies calls or stores user tokens.