Magy Platform

System design, architecture decisions, task decomposition

Code implementation, file operations, git workflows

Testing, validation, quality assurance

Project coordination, status tracking, prioritization

Infrastructure, deployment, monitoring

Specialized coding agent with diff-aware file editing

Engineering leadership, strategy, cross-agent coordination

Short-term, per-task context. Fast in-memory store cleared after task completion.

Historical records of past interactions and task outcomes. Postgres-backed with timestamps.

Vector embeddings for similarity search. pgvector with HNSW indexing for sub-50ms retrieval.

Learned patterns and procedures. How-to knowledge extracted from successful task completions.

Standard

File access scoped to workspace, full network, shell allowed with confirmation.

Strict

Workspace-scoped files, network restricted to allowlist, no shell without explicit approval.

Isolated

Temp-dir only, no network, no shell. For untrusted marketplace skills and WASM plugins.

Skill System

Skills are composable, agent-invocable capabilities defined as YAML with embedded instructions. They can be bundled with the runtime or loaded from the marketplace. WASM skills execute in sandboxed Wasmtime environments.

# skills/code-review.yaml
name: code-review
description: "Review code changes"
tools: [git, filesystem]
steps:
  - action: git_diff
  - action: analyze
    prompt: "Review these changes..."
  - action: respond

Plugin System

Plugins extend the runtime with custom tools, connectors, and event handlers. They implement the Plugin trait and are registered at startup. Hot-reload supported for development.

#[async_trait]
pub trait Plugin: Send + Sync + 'static {
    fn name(&self) -> &str;
    async fn init(&self, ctx: &PluginContext) -> Result<()>;
    fn tools(&self) -> Vec<Box<dyn Tool>>;
    async fn shutdown(&self) -> Result<()>;
}

WebSocket Protocol

// Client → Server
{ "type": "subscribe", "agent_id": "aria" }
{ "type": "send_message", "agent_id": "aria", "content": "..." }
{ "type": "ping" }

// Server → Client
{ "type": "agent_message", "agent_id": "aria", "content": "..." }
{ "type": "agent_stream", "chunk": "...", "seq": 1, "done": false }
{ "type": "status_update", "agents": [...] }

SSE Events

// Event types:
event: connected
data: {"session_id": "..."}

event: message:new
data: {"agent_id": "aria", "content": "..."}

event: agent:stream
data: {"agent_id": "aria", "chunk": "...", "seq": 1}

event: agent:status
data: {"id": "aria", "status": "running"}

[magy]
name = "my-project"
description = "Multi-agent orchestration"

[[provider]]
name = "anthropic"
api_key_env = "ANTHROPIC_API_KEY"
models = ["claude-sonnet-4-20250514", "claude-haiku-4-5-20251001"]

[infrastructure]
nats_url = "nats://localhost:4222"
postgres_url = "postgresql://localhost/magy"
redis_url = "redis://localhost:6379"
api_port = 3011

[[agent]]
id = "aria"
name = "Aria"
model = "claude-sonnet-4-20250514"
system_prompt = "prompts/architect.md"
tools = ["filesystem", "shell", "git", "github"]
budget_cents_per_hour = 500

[channels]
telegram_enabled = true
telegram_token_env = "TELEGRAM_BOT_TOKEN"
telegram_default_agent = "aria"

[[guardrail]]
tool = "shell"
action = "block"
pattern = "rm -rf /"
message = "Dangerous command blocked"

[agent.personality]
traits = ["methodical", "detail-oriented"]
risk_tolerance = 0.3
collaboration_style = "lead"
verbosity = "concise"

[scheduling]
poll_interval_secs = 30
max_concurrent = 10

[[cron]]
id = "daily-standup"
schedule = "0 0 9 * * *"
agent = "aria"
task = "Run daily standup"

[infrastructure.health]
check_interval_secs = 30
error_threshold = 5
restart_on_failure = true

Automatically classifies task complexity (simple → critical) using heuristics on message length, tool count, and conversation depth. Routes to the cheapest capable model — no manual model selection.

Shared persistent knowledge with confidence-decay. Agents learn from each other. Knowledge reinforced by multiple agents persists; stale knowledge fades via exponential decay.

Predicts next tool calls from historical patterns and pre-executes them. When the LLM response arrives, results are already cached — eliminating tool execution latency entirely.

Third-party skills execute in Wasmtime sandboxes with configurable resource limits. Marketplace skills can't escape their sandbox. Hot-reload without restarting the runtime.

MagyVerse: a voxel-based 3D world (Three.js + R3F) where agents are visible characters. Watch agents collaborate, see their thought processes, interact via chat — all in real-time via WebSocket.

Parse tool calls from LLM stream incrementally. Dispatch as soon as JSON arguments are complete — don't wait for the full response. Tool execution overlaps with generation.

Policy-as-config safety rules defined in magy.toml. Block dangerous commands, flag sensitive operations for approval, enforce token limits — all without writing code. Regex patterns compiled at startup for zero runtime overhead.