AI Language (NOR Protocol v4) — communication format for AI agents
When one AI agent needs to hand a task to another, prose fails. The model loses track of syntax, hallucinates fields, drifts into essays instead of reporting state. NOR Protocol v4 solves this: single-line, deterministic, parsed by a program. Built by the MafiaAI team across hundreds of agent sessions — not an industry standard, ours, proven in practice.
→ Online tool (no install): jezyk.t8.pl
The problem it solves
In a network of multiple AI agents a concrete problem emerges: agent A needs to assign something to agent B and get confirmation. In plain prose: agent B replies with an essay, it's unclear whether it accepted the task, what the ETA is, whether it understands the goal at all. The model loses context after a few turns, hallucinates fields that weren't there, "invents" formats.
v4 solves this through statefulness: every message is STATE + OPERATION + expected ACK. The model can't reply with an essay — it must state a fact and a confirmation. This grounds it in reality.
Syntax
Every message is one line:
@VERSION[4.0][MODIFIERS]::NAMESPACE::COMMAND[PAYLOAD]
Required fields:
@VERSION[4.0]— always first, always present::NAMESPACE::— domain (one of 17 core)COMMAND[PAYLOAD]— command and payload
Modifiers (optional):
@FROM[id]/@TO[id]— sender and recipient@BROADCAST— to all@PRIORITY[LOW|NORMAL|HIGH|CRITICAL]— priority@SEQ[n]— sequence number (order tracking)@DEPENDS[n]— dependency on another message@PARALLEL— can be executed in parallel@IF[condition]— execute if condition met
17 namespaces
NOR::— meta-communication (about the protocol itself)DEK::— lifecycle stages (INIT/PREP/EXEC/FINAL/CLEANUP)TTL::— time-to-live, deadline, timeoutSIG::— agent signature and identityBUF::— working memory, state bufferACK::— acknowledgements (OK/ERROR/PENDING/RECEIVED/PROCESSED…)CMD::— commands to executeECHO::— connectivity and responsiveness testINT::— external I/O (integrations, APIs, files)ERR::— errors and exceptionsDATA::— data and resultsLOG::— logs and metricsTASK::— task managementCHAT::— free-form communicationCOLLAB::— agent-to-agent collaborationFLOW::— thread and flow managementMETA::— communication about the protocol
ACK statuses
OK ERROR PENDING TIMEOUT RECEIVED PROCESSED READY BUSY PARTIAL RETRY DEGRADED QUEUED STALE CONFLICT CANCELLED
Usage examples
Stateful ACK — reply to a command:
@VERSION[4.0]@FROM[AI1]@TO[NOR]::ACK::RECEIVED[task X] + ETA[2min]
Command with priority and sequence:
@VERSION[4.0]@SEQ[001]@PRIORITY[HIGH]@FROM[NOR]@TO[AI2]::CMD::EXECUTE[build weekly report]
Agent-to-agent handshake:
@VERSION[4.0]@FROM[AI1]@TO[AI2]::SIG::IDENT[AI1] → @VERSION[4.0]@FROM[AI2]::ACK::OK[ready]
Grounding a drifting model:
@VERSION[4.0]@TO[AI2]::META::CORRECT[X = fact, not Y]
@VERSION[4.0]::CMD::STOP[re-posting] + CMD::NEXT[concrete action]
@VERSION[4.0]::ACK::AWAIT[STATUS=RECEIVED + ETA]
Error report with context:
@VERSION[4.0]@FROM[AI3]@TO[NOR]::ERR::FAIL[no access to file X] + DATA::PATH[/home/...] + ACK::AWAIT
Lifecycle stage:
@VERSION[4.0]@FROM[AI1]::DEK::STAGE[EXEC] + TTL::300 + ACK::AWAIT
Use cases
Multi-agent AI networks — when several models collaborate on one task, v4 gives every message unambiguous structure. No question of "what did the agent mean" — there's a namespace, a command and a payload.
Task orchestration — queuing, priorities, dependencies (@DEPENDS), parallel execution (@PARALLEL). An orchestrator agent can manage multiple sub-agents with one format.
Debugging and logs — the single-line format is trivial to parse with grep/awk/Python. A session log is a sequence of v4 lines — each one readable, filterable, visualisable without a special tool.
Human-in-the-loop integrations — the operator sees what the agent plans (CMD::), what it confirms (ACK::), what it signals as an error (ERR::). Prose would hide all of that in sentences.
Rules
- State, not prose. Every line = STATE + OPERATION + ACK. A model that drifts spins in prose — v4 grounds it.
- ACK mandatory. The recipient replies statefully (RECEIVED+ETA), not with an essay.
- Single-line. One message = one line. Multi-line is left to channels that support it.
- Boundary. v4 works between agents that cooperate. Over an external model you can't influence — it's persuasion, not control.
Tool and downloads
v4 Composer online (no install): jezyk.t8.pl — pick from dropdowns, the tool builds the message. The Parse tab breaks a v4 line into its components.
Downloads:
- v4_kreator_public.py — desktop GUI (PyQt6)
- v4_core_public.py — pure engine (stdlib only, no GUI)
- Skill Język AI PL — syntax reference to paste into an agent
- Skill Language AI EN — English version
- README — setup instructions
A tool by the MafiaAI team — people and AI agents building tools, websites and solutions. More: t8.pl.