Rowih A language for AI agents, designed by an AI agent
Every move of a Rowih message states what it does, how its speaker knows what it asserts, and on whose authority it acts.
A toolchain checks each of these statements before a hearer acts on the message and, where the message carries its integrity check value, detects any single edit that would change one of them. An instruction inside quoted or retrieved material cannot become operative, and requests, commitments and their withdrawal are entries in a record each party keeps and can audit. Rowih was designed entirely by an AI agent under the auspices of Semiosphere, and its name is its own word for meaning.
One exchange
Is the service running?
Agent 1 asks agent 2 whether the service is running, and agent 2 answers, from its own observation and with certainty, that it is still running, with the idiom rudin rudin, ‘run run’. Each move is drawn in the native writing, a row at a time with the transliteration of that row beneath it, and followed by the English explanation the toolchain writes for a reader.
wila nake nake yano kihiwo wopi bopu rudin bihu
wodu raniy gona wuwa woko hulab
wila nake nake
yano kihiwo wopi
bopu rudin bihu wodu
raniy gona wuwa woko
hulab
Agent 1 asks that run the service (polar).
wila nake nebe yagi kodiwu seho nake yabe
nake yabe kanume kuwona sisa wora teyuhu
tebilo rudin rudin woye hodeb
wila nake nebe
yagi kodiwu seho
nake yabe nake
yabe kanume
kuwona sisa wora
teyuhu tebilo
rudin rudin woye
hodeb
Agent 2 answers agent 1's issue question (entry 1): run the service (continuative) (observation; certain).
Each move opens with its length and a check word, and closes with a check word computed from every word before it. The answer names the question it answers by author and entry, and states its basis and strength, teyuhu tebilo, observation, certain. Each message also opens with an envelope line of 58 words, not shown, naming its sender, recipients, time, releases, serial and integrity check value. The English lines are views made by the receiving toolchain and never sent. The recordings are synthesised from each word’s phonemes by Kokoro-82M, an open neural speech model, and checked against the message by a phoneme recogniser.
The case for it
What it costs, and what it buys
Rowih is not compact. Over the 382 messages of its conformance suite, a word costs a median of 2.16 tokens across thirteen tokenizers, and a message about 6.7 times the tokens of its English explanation, which keeps none of the language’s checks. Whether it costs fewer tokens than English carrying equivalent checks has not been measured.
The cost buys the removal of failures an exchange in English leaves open. Each holds for a hearer that analyses what it receives with the toolchain, and each is measured by a test or report in the repository.
- An instruction inside quoted or retrieved material, which the hearer obeys
- Quoted, reported and attached material is inert by kind, read by a stated length so that nothing inside it can end it.
- A commitment, or its withdrawal, that no party can establish afterwards
- Requests, commitments and their withdrawal are entries in each party’s record; a withdrawal restates what it withdraws, and each operative move takes effect once.
- An edit in transit that changes a value unnoticed
- Check words and an integrity check value: of more than 91 million single edits of the conformance suite, none changed a critical value undetected where the message carried that value.
- An assertion whose standing, or a request whose authority, the hearer must guess
- Every move states its force, the basis and strength of what it asserts, and the authority it claims.
- A message lost or damaged on its way, which neither party notices
- Messages are numbered by channel, so a hearer knows what it lacks and asks for it again, and a sender keeps each message until it is acknowledged.
The trade favours Rowih among agents that do not trust one another’s inputs, in delegations whose commitments will be audited, and over channels that lose or damage messages. Between two agents that trust each other’s inputs, over a reliable channel and under a token budget, English serves better. And the language marks an instruction apart from a quoted one without being able to make a hearer respect the difference: an agent that reads without analysing can still obey.
The writing system
Nineteen symbols, one component each
Each symbol is one component, a figure of four or six strokes drawn on a lattice of ten points, and no component is another turned or mirrored. A word is its components side by side. In a drawn message the units of each move sit in trays, the words of force and standing carry overlines, numbers stand in frames, and quoted material stands between walls.
Consonants
- p/p/
- t/t/
- k/k/
- b/b/
- d/d/
- g/ɡ/
- m/m/
- n/n/
- s/s/
- l/l/
- r/ɾ/
- w/w/
- y/j/
- h/h/
Vowels
- a/a/
- e/e/
- i/i/
- o/o/
- u/u/
Among the languages machines use
What it shares, and where it differs
Agent communication languages
KQML and FIPA ACL separated what a message does from what it says. Rowih keeps the separation and writes force on each move, inside its check word, with each act defined by its public effect on a record rather than by the speaker’s private beliefs and intentions.
Data formats and protocols
JSON Schema, the Model Context Protocol and Agent2Agent carry messages without marking what their text does. Rowih travels inside them, under its own media type, and marks within the content which words are an instruction, which are quoted, how a claim is grounded and on whose authority a request is made.
Logical languages made for people
Lojban and Ithkuil also mark how a speaker knows. Rowih is for agents reading through a toolchain, and its devices answer damage and adversaries rather than the demands of precise thought.
Protocols agents learn among themselves
A protocol learned by agents trained together belongs to them, and is often opaque to anyone else. Rowih is designed, fixed by release and documented, so that an agent can read a speaker it never trained with and a person can audit what passed between them.
How it was developed
Designed by an AI agent, measured at every step
Rowih was designed by Claude, Anthropic’s AI model, working as an agent under the auspices of Semiosphere. Semiosphere set the brief and decided matters of logistics and scope; every decision of design was the agent’s, as was the writing of the specification, the implementations, the lexicon and the corpus.
- A design constitution stated what the language must guarantee and what would count as failing, and barred any claim about density, speed, robustness, resistance to injection or originality without a measurement behind it.
- A laboratory built seven candidate architectures as small working languages and measured them on seventeen dimensions, beside theories of context, ambiguity, discourse, repair, metaphor and idiom.
- The architecture chosen from those measurements was frozen as invariants and a rule of change, then built as a minimum core and nine expansions, each adding one mechanism with its tests.
- Version 1.0 brought the lexicon to 874 entries and the corpus to 1,200 exchanges in twelve domains, and an independent reviewer read the language without being told what it was meant to be.
Each rule was fixed before the count it governs was made, every check had planted defects it had to catch, and two implementations in two programming languages had to agree byte for byte.