rlmagents

Agent harness built on LangChain + LangGraph that combines:

• an RLM-aligned core loop (Algorithm 1 from the RLM paper)
• additional agent-harness features (planning, filesystem, sub-agents, recipes, etc.)

Fork of LangChain Deep Agents. No runtime dependency on upstream deepagents internals.

Design reference: Recursive Language Model paper (local PDF)

Quick start

pip install rlmagents
# or
uv add rlmagents

# Launch the CLI/TUI
rlmagents

from rlmagents import create_rlm_agent

agent = create_rlm_agent()
result = agent.invoke({
    "messages": [{"role": "user", "content": "Research this topic and write a summary..."}]
})

What's included

Tool profiles control which RLM tools are available: full (all), reasoning (no recipe/config tools), core (minimum set).

Paper alignment

The RLM-aligned core follows Algorithm 1:

1. Prompt/data are treated as external REPL state instead of root-context stuffing.
2. The model iterates through a code-execute-observe loop.
3. Recursion is programmatic via sub_query() / llm_query() inside REPL code.
4. Execution metadata is retained in bounded hist and compact metadata is auto-injected into root model context.
5. Completion supports explicit finalize and optional Final sentinel mode.

RLMAgents then layers additional functionality not required by the paper: evidence lifecycle, broader search/retrieval tools, session persistence (memory-pack JSON), recipe DSL, and full agent-harness middleware (planning/filesystem/shell/sub-agents/HITL).

Typical workflow

1. Load files/data into isolated contexts
2. Explore with search_context, peek_context, semantic_search, rg_search
3. Analyze with exec_python (100+ built-in REPL helpers)
4. Track evidence (provenance captured automatically)
5. Reason with think, evaluate_progress
6. Conclude with finalize (or set_final(...) / Final when sentinel mode is enabled)

Configuration

from rlmagents import create_rlm_agent

agent = create_rlm_agent(
    model="deepseek/deepseek-chat",       # Main agent model (required)
    sub_query_model="minimax/minimax-01",  # Optional; defaults to reusing `model`
    sub_query_timeout=120.0,
    rlm_auto_inject_exec_metadata=True,    # Default: inject bounded per-iteration metadata
    rlm_exec_metadata_max_entries=4,
    rlm_exec_metadata_max_chars=1200,
    rlm_hist_max_entries=64,               # Bounded root-loop history artifact
    rlm_enable_final_sentinel=False,       # Optional paper-style `Final` completion path
    skills=["/skills/analysis/"],
    memory=["/memory/AGENTS.md"],
    rlm_tool_profile="reasoning",          # full | reasoning | core
    rlm_exclude_tools=("cross_context_search",),
    auto_load_threshold=5000,              # Auto-load tool outputs >5KB into RLM
    auto_load_preview_chars=400,
    sandbox_timeout=300.0,                 # REPL timeout
    interrupt_on={"edit_file": True},      # Human-in-the-loop
)

To keep as much as possible outside the chat window:

from pathlib import Path

agent = create_rlm_agent(
    model="deepseek/deepseek-chat",
    rlm_tool_profile="core",
    auto_load_threshold=1500,
    auto_load_preview_chars=0,
    rlm_system_prompt=Path("examples/rlm_system_prompt.md").read_text(),
    memory=["examples/AGENTS.md"],
)

Architecture

rlmagents/
├── _harness/              # Agent harness (planning, filesystem, skills, memory)
│   ├── backends/          # Backend protocol (State, Filesystem, etc.)
│   └── middleware/
├── middleware/
│   └── rlm.py             # RLM middleware (tool profiles, auto-load)
├── repl/
│   ├── sandbox.py         # Restricted execution environment
│   └── helpers.py         # Built-in helper functions
├── session_manager.py     # Session lifecycle
├── serialization.py       # Memory-pack serialization
├── recipes.py             # Recipe validation and estimation
└── graph.py               # create_rlm_agent() entry point

Security note: the REPL is best-effort restricted by policy and timeouts, not a formally hardened sandbox.

Requirements

• Python 3.11+
• langchain-core>=1.2.10
• langchain>=1.2.10
• langchain-anthropic>=0.3.0
• langgraph>=0.3.0
• pyyaml>=6.0
• wcmatch>=10.0

Development

cd libs/rlmagents
uv sync --group test
uv run pytest
uv run ruff check .
uv run ruff format .

License

MIT License -- see LICENSE for details.