v2.0.0 - The Steward Architecture

The Steward Architecture

v2.0.0 introduces agents as stewards of long-lived systems — autonomous agents that own a domain, maintain it over time, react to change, coordinate with other stewards, and survive crashes.

Highlights

• Persistent Beliefs: @persistent annotation for agent fields that survive restarts
• Supervision Trees: OTP-style restart strategies (OneForOne, OneForAll, RestForOne)
• Session Types: Compile-time verified communication protocols
• Effect Handlers: Per-agent LLM configuration
• Observability: trace() and span blocks with OTLP export
• Built-in Tools: HTTP, Database, Filesystem, Shell with capability-based access

Example

agent Counter {
    @persistent count: Int

    on waking {
        trace("Restored at {self.count.get()}");
    }

    on start {
        self.count.set(self.count.get() + 1);
        yield(self.count.get());
    }
}

supervisor App {
    strategy: OneForOne
    children {
        Counter { restart: Permanent, count: 0 }
    }
}

run App;

Performance

Metric	Target	Measured
Startup (check)	< 100ms	~10ms
Checkpoint latency	< 10ms	< 1ms
Restart latency	< 50ms	~1ms

Documentation

• Persistent Beliefs Guide
• Supervision Trees Guide
• Session Types Guide
• The Steward Pattern

See the full changelog for details.