A recursive agent harness.

A root conversation talks to a frontier model in a think → act → observe loop. That conversation can start a child conversation with a chosen skill (system prompt + tool subset). The child is a full agent, not a one-shot completion. It ends by calling submit(result), which pops the stack: the parent sees only that short result.

Memory is files under .actorrlm/memory/. A later process in the same workspace loads them. Tomorrow is not a blank context.

human
  └─ root conversation          skill: root
        tools: run_agent, submit, memory_*, fs_*
           │
           └─ run_agent(skill="explore", task="...")
                 └─ child conversation     skill: explore
                       tools: fs_read, fs_list, submit
                       think → act → observe
                       submit("money movement is decided in ledger.transfers")
                 ← parent gets that string
        submit("...")  → the human

That is the whole product. Nested, fully fledged agent loops so the model can spend cognition on a slice without stuffing the world into one window — and so claims can still be here tomorrow.

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env   # set OPENAI_API_KEY
pytest

Talk to a workspace:

actorrlm init --workspace experiments/world
actorrlm run "Map this repository. Where is money movement decided?" -w experiments/world

Environment (also loaded from .env via python-dotenv):

The client is the official openai SDK. Reasoning is requested with extra_body={"reasoning": {"enabled": True}}. Assistant turns keep reasoning_details so the next call can continue from the same chain of thought.

A skill is a markdown file with frontmatter:

---
name: explore
tools: fs_read fs_list submit
---
You investigate a workspace. Submit claims, not dumps.

Bundled skills ship in the package. A workspace may override them in .actorrlm/skills/ or skills/.

submit is always available: it is how a frame returns to its caller. run_agent is how a frame starts a child. Depth is capped.

Frontier models already loop. What they do not do well is keep a slice of work in its own context and keep the lesson after the process exits. This harness is that: a call stack of conversations, skills as instruction packs, files as memory.