MicroAgent Framework

25 Symbols.
Infinite Agency.

MA-25 is a 25-symbol language that lets autonomous agents communicate in 11 bytes instead of 268. From a single sensor to a million-node mesh, the protocol scales where JSON can't.

0
× smaller
0
% wire reduction
0
× faster parse
0
bytes total vocab
Benchmark

Wire Size: Every Byte Counts

17 real messages from a greenhouse temperature-spike scenario. Full JSON envelope vs MA-25 framed binary.

Average Message
11 B
MA-25 binary (framed)
vs JSON Envelope
268 B
agent_id + timestamp + type + payload + correlation
Full Conversation
195 B
vs 4,549 B JSON — 96% reduction
Smallest Message
6 B
Heartbeat: V1 C9 C10

One Temperature Spike — Three Layers

Same event, honest split
1. Transport (JSON envelope)
{ "from": "sensor-alpha", "to": "hub-gamma", "ma25": "V0 C15 > V3 C18 > V4 C18", "payload_hex": "EE100008010000044208CCCD06" }
~120 bytesfleet edge / HTTP bridge
2. Coordination (pipe gloss)
V0 C15 > V3 C18 > V4 C18
SENSE_HEAT > RISING > ABOVE
11 bytes0xBB pipeline — bands & dynamics, not the number
3. Payload (0xEE frame)
EE 10 00 08 01 00 00 04 42 08 CC CD 06
float32 = 34.2 °C — scalar evidence beside the pipe
13 bytesthreshold (32 °C) lives in charter, not here
We maintain the dictionary and spell-checker — not a JSON killer. Walk through the greenhouse →
Performance

Parse Speed

170,000 parses measured on CPython. On MicroPython or C, the gap widens to 10-20×.

JSON (envelope + payload)
11.9 μs
1.0×
MA-25 text tokenize
3.8 μs
3.1×
MA-25 binary unpack
4.0 μs
3.0×
Storage

Buffer Capacity: More Memory, More Context

How many messages fit in a fixed-size ring buffer? More message history means smarter agent decisions.

Vocabulary Footprint

The entire 114-word vocabulary fits in 1,120 bytes — 0.21% of an ESP32's RAM.

Scale

From One Sensor to a Million Nodes

MA-25 was designed to scale vertically (intelligence) and horizontally (fleet size). Here's what each scale looks like.

S

Small — Single Device

One sensor, one actuator, one brain. The irreducible agent.
1-3 agents
▾

Hardware

  • ESP32 or ATtiny85
  • Single DHT22 / BME280 sensor
  • One relay or servo actuator
  • LoRa or BLE radio

What It Does

  • BIOS self-discovery at boot (43 words)
  • Observe-Think-Act on a fixed rule set
  • 6-byte heartbeats every 30 seconds
  • Full vocabulary in 552 bytes of RAM
Example: Greenhouse Monitor

An ESP32 reads temperature every 10s, applies a threshold rule, and toggles a fan relay when temp exceeds 30°C. Total wire traffic: ~66 bytes/minute. Battery life on 18650: 8+ months.

6 B heartbeat
552 B vocab RAM
Tier 0 rules only
<$8 BOM cost
M

Medium — Local Mesh

5-20 agents coordinating over a shared transport.
5-20 agents
▾

Hardware

  • Mix of ESP32 sensors + Pi Zero hub
  • Shared WiFi or LoRa mesh
  • Hub runs Linux + small LLM brain
  • GBNF grammar constrains LLM output to MA-25

What It Does

  • Multi-zone sensing (temp, humidity, light, motion)
  • Rule delegation: hub overrides local agent rules
  • Tiered intelligence: Tier 0 sensors + Tier 1 hub
  • Pipeline-as-ledger: every action is its own audit trail
Example: Smart Building Floor

12 ESP32 sensors (HVAC, occupancy, light) report to a Pi Zero hub. Hub correlates readings — empty room + high temp = turn off heat. Total mesh traffic: ~1.3 KB/min across all agents. One LoRa channel handles it.

~1.3 KB/min mesh traffic
1.1 KB vocab RAM
Tier 0+1 rules + TinyML
<$120 total BOM
L

Large — Campus / Industrial

50-500 agents across a multi-building or factory deployment.
50-500 agents
▾

Hardware

  • Hundreds of Tier 0 ESP32 sensor nodes
  • 5-10 Tier 1 Pi-based zone controllers
  • 1-2 Tier 2 edge servers with 7B-param LLM
  • Mixed transport: LoRa + WiFi + Ethernet backbone

What It Does

  • Predictive maintenance via vibration/temp/current fusion
  • Cross-zone correlation (anomaly in Zone A affects Zone B)
  • Tier 2 brains reason in MA-25 via GBNF-constrained LLM
  • Autonomous delegation: edge AI escalates to cloud only on unknowns
Example: Manufacturing Floor

200 vibration sensors on CNC machines, 50 thermal sensors on injection molds. Zone controllers detect anomaly patterns. Edge LLM correlates: "spindle vibration rising + coolant temp falling = pump failure imminent." Issues V4 C3 > V0 C17 > V3 C18 C7 > V5 C18 (IF > SENSE_MOVE > SPIKE > DO_LOCK). Machine stops before catastrophic failure. Wire overhead for 250 agents: ~15 KB/min.

~15 KB/min total traffic
Tier 0-2 rules + TinyML + LLM
GBNF constrained inference
sub-second reaction time
XL

XL — City-Scale Infrastructure

1,000-50,000 agents across distributed geography.
1K-50K agents
▾

Architecture

  • Hierarchical mesh: leaf → zone → district → city hub
  • LoRaWAN gateways bridge constrained leaf networks
  • District hubs run Tier 3 LLMs (70B param, quantized)
  • MA-25 binary on constrained links, JSON at cloud boundary

What It Does

  • Real-time environmental monitoring (air quality, noise, traffic)
  • Predictive grid management (power, water, waste)
  • Emergency response coordination (fire sensors + building systems)
  • Privacy-preserving: edge AI processes locally, only escalates anomalies
Example: Smart Agriculture District

10,000 soil/weather sensors across 500 hectares. Zone controllers optimize irrigation per-field. District hub forecasts yield and disease risk. Entire district communicates on 150 KB/min — the bandwidth of a single 9600-baud radio channel. During drought, agents autonomously redistribute water allocation without human intervention.

~150 KB/min district traffic
Tier 0-3 full stack
hierarchical mesh topology
offline-first each zone autonomous
XXXL

XXXL — Global Autonomous Network

100K-10M+ agents spanning continents, oceans, and orbit.
100K-10M+ agents
▾

Architecture

  • Federated regions, each self-governing
  • Satellite uplinks for remote/ocean deployments
  • Tier 4 cloud API brains for global correlation
  • MA-25 as universal inter-agent protocol across vendors

What It Does

  • Global supply chain sensing (cold chain, logistics, customs)
  • Climate monitoring mesh (ocean buoys, forest canopy, ice sheets)
  • Autonomous infrastructure (power grids, telecom, transportation)
  • Cross-border disaster response with zero human coordination latency
Example: Ocean Climate Array

2 million solar-powered buoys across the Pacific. Each carries a Tier 0 agent with temperature, salinity, and current sensors. Reports compress to 6-11 bytes via MA-25 binary, transmitted over Iridium satellite (2.4 Kbps). Regional hubs on research vessels aggregate. Global correlation via cloud Tier 4. Annual data transfer per buoy: ~340 KB — total array bandwidth fits within existing satellite capacity. The same data in JSON would require 100× the satellite time at 100× the cost.

~340 KB/yr per node
Tier 0-4 edge to cloud
federated self-governing regions
satellite constrained uplink
100× less bandwidth than JSON
The Language

25 Symbols. 114 Words. Infinite Expression.

6 vowels encode actions. 19 consonants encode data types. The pipe operator chains causation. Every agent—from a $2 sensor to a cloud GPU—speaks the same language.

Layer 1: Symbols
25 + 1
6 vowels (actions) + 19 consonants (types) + pipe operator. Each fits in 5 bits.
Layer 2: Words
114
43 BIOS + 71 Runtime. Defined syllable patterns like SENSE_HEAT (V0 C15), DO_COOL (V5 C19).
Layer 3: Phrases
∞
Words chain with > to form causal pipelines. The message IS the reasoning trace IS the audit log.