Running on Raspberry Pi 1 (512MB RAM)

Self-Evolving
Multi-Agent Framework

Agents that modify their own behavior at runtime through a declarative genome.
Zero downtime. 3MB binary. 100% Autonomous.

play_circle Interactive Demo code View Code
Paper by Alex Chen & Bowen Li • Independent Researchers

The Static Agent Bottleneck

Traditional agents (AutoGPT, LangChain) are fundamentally static. Once deployed, their capabilities are frozen in the binary. To learn a new skill or adapt to a new API, engineers must rewrite code, recompile, and redeploy.

block Current Limitations

  • Monolithic binaries (>200MB)
  • Cannot run on Edge devices
  • Manual updates required

The EVOCLAW Solution

EVOCLAW treats agent capabilities as a Genome. The runtime engine is fixed, but the behavior is defined in a hot-swappable genome.toml file. Agents evolve by mutating this file.

check_circle Genome Advantages

  • Hot-reload without restart
  • 3.0 MB Binary Size
  • Evolutionary Optimization

Interactive Genome Engine

Modify the declarative genome below to see how the agent adapts in real-time.

genome.toml
[genome]
id = "pi1-edge-agent"
generation = 42
mutation_rate = 0.05

[skills.system_monitor]
enabled = true
interval_sec = 30

[skills.price_monitor]
enabled = true
symbols = ["BTC", "ETH"]

[skills.gpio]
enabled = false
pins = [17, 27]

memory Runtime Status

RAM Usage (RSS)
3.2 MB
Active Skills
2
CPU Load (Sim)
2.1%
Binary Size
3.0 MB

Active Skills

monitor_heart System Monitor
30s interval
currency_bitcoin Price Monitor
Default
[INFO] Genome loaded. Generation 42.
[INFO] Skill 'system_monitor' initialized.
[INFO] Skill 'price_monitor' initialized.
[MQTT] Connected to broker.

Multi-Tier Architecture

Layer 1: Orchestrator

Central Management (Go)

  • check Manages fleet of agents
  • check Mutates genomes
  • check Validates constraints

Layer 2: Message Bus

Async Communication

  • hub Topic: agents/{id}/genome
  • hub Topic: agents/{id}/report
  • hub QoS 2 (Exactly Once)

Layer 3: Edge Agent

Runtime Engine (Rust)

  • memory Static Binary (3MB)
  • bolt Hot-Reload Engine
  • hardware GPIO / Sensor Access

Deployment Flexibility

From Raspberry Pi 1 (Edge) to Docker Containers (Server) to E2B Sandboxes (Cloud). One genome format for all.

Performance & Evaluation

RAM Footprint Comparison

* EVOCLAW running 3 concurrent skills on Raspberry Pi 1.

72-Hour Stability Test (Raspberry Pi 1)

0
Crashes
100%
Uptime
13,807
MQTT Messages
<100KB
Memory Leak
warning Resilience

The agent successfully recovered from 12 transient network failures during crypto price fetching using exponential backoff.

Metric Edge (Pi 1) Server Cloud (E2B)
Binary Size 3.0 MB 7.2 MB N/A
RSS Memory (3 Skills) 3.2 MB 12.4 MB 45 MB
Hot-Reload Time 108 ms 35 ms 60 ms
CPU Usage 2.1% 0.3% 0.1%