An executable language for human–AI and AI–AI communication.
Programs are spatial-symbolic compositions: objects, rooms, flows, colours and their order. One set of operators acts on any host, from AI agents to teams and organisations. A cross-domain engine, the White Dragon, redesigns the host it runs on. Consent and harm checks are part of the compiler.
See the measured resultsTry the editorPre-registered tests on real LLM agents (more than 23,000 model calls). 41 predictions on real hosts: 25 passed, 16 failed. The failures are on the results page, next to the passes.
RSOS State Runtime: 2 model calls instead of 42
For multi-agent workflows that edit shared state. The model decides what should change; the RSOS runtime executes it, checks every agent's copy and heals drift. A whole task fits in one model call, because the operations name keys and the runtime computes every value.
Drop it in: Python or MCPReplicate it, under $1| Best setup for the same agents | RSOS | |
|---|---|---|
| Model calls, 14-step task | 42–43 | 2.2–2.4 |
| Cost per task | 1× | 0.27–0.35× |
| Time per task | ~100 s | 12 s |
| Tasks fully correct | 31/32 | 32/32 |
# Claude Code, or any MCP client
pip install rsoslang
claude mcp add rsos -- rsoslang mcp
What a program looks like
A mental model placed on a real object. The work table is a robot; each part carries one composed model; a symbol anchors it and opens it on demand. The same lines compile for a person's practice or for an AI agent's workflow.
This is examples/robot_head.rsl from the repository. It compiles with
python -m rsoslang check to 23 operations and 0 errors.
HOUSE mind FOR person
ROOM eagle PERSPECTIVE above IN mind
OBJECT robot KIND work_table IN eagle
PART robot.head KIND decision_model
MODEL robot.head = associated + confidence
+ analogical_reasoning IN eagle FOR decision
MODEL robot.hand = top_view + dissociated
+ future_pacing + action IN eagle FOR action
OBJECT couch KIND furniture IN eagle REPRESENTS langgraph
DEDICATE couch TO hypothesis_testing
FORM couch DIRECTIONS forward -> hypothesis
; backward -> contradiction ; up -> framing
ANCHOR eagle_calculator TO robot.head
FUTUREPACE eagle_calculator SCENARIOS 3
OPEN eagle_calculator FROM kinesthetic SCOPE whole
Four parts, one language
Self-healing communication
Agents share a book of meanings. Each message carries a green dot, a check of the state it refers to, so drift is caught and the changed part is healed from the book, with no model call.
A1, A3, A3b →The White Dragon engine
It breaks a host into parts, picks a mechanism from another domain for each, and sets the depth, the order and the predicted gain. Tested against blind and weighted choice.
A2 →Mental-model construction
Reasoning and decisions are built as a house: rooms are states, objects are models, each part of an object carries one model. People run and edit them; they do not just recall them.
Programme →Governance in the compiler
Operations on people need their consent (E018), standing consent needs a declared network with laws (E021), and the Spine blocks predicted harm (E019), all at compile time.
Responsible use →The claim, worded to its evidence
In pre-registered tests on real LLM agents, RSOS carried every message losslessly in 0.141× the BPE channel tokens of a Slipstream-style anchor protocol with zero model calls (0.106× with a local predictor), kept every copy of the shared state true, and repaired every injected drift without a model call. When agents computed from their copies, the anchor protocol got 0 of 16 tasks right; RSOS got 15–16. Written as programs the runtime executes, a 14-step task took 2 model calls instead of 42.
Check values and restore-from-journal are old ideas in databases and file systems. To our knowledge, RSOS is the first agent communication language in which detection and in-place healing are part of the protocol, and the first whose protocol was designed by a cross-domain engine tested against blind alternatives.
Missed: 10× on the model bill (A5 reached 2.8–3.7× at API prices; the limit is now the model's thinking tokens). The 5× channel target missed in A3b was reached in A3c. Next: A5b, thinking control applied equally to RSOS and the baseline.