A Multi-Agent Orchestrator with Token Budgets and a Time Machine

What it does

This self-hosted platform routes user tasks through execution strategies—simple, DAG, ReAct, research, swarm—selected by complexity scoring. It exposes a REST API, an OpenAI-compatible proxy, a Python SDK, and a Tauri desktop app, backed by Go (gateway/orchestrator), Rust (agent core), and Python (LLM adapters). Temporal workflows provide execution durability, while Prometheus and OpenTelemetry handle observability.

The interesting bit

Execution is treated as infrastructure rather than scripting: code runs inside a WASI sandbox, actions are gated by OPA policies and human-approval WebSockets, and per-task token budgets trigger automatic model fallback. Time-travel debugging lets operators rewind and replay past executions incrementally—an admission that production agents fail, so you need a rewind button.

Key highlights

• Complexity-based routing picks from eight strategies, including tiered research models that the project claims can cut costs by 50–70%.
• Hard token budgets per task and agent, with automatic failover across 10+ providers (OpenAI, Anthropic, Google, DeepSeek, xAI, Ollama, etc.).
• WASI sandbox for code execution, plus OPA policies and multi-tenant isolation.
• Human-in-the-loop approval gates routed via WebSocket to connected daemon clients.
• Temporal-backed workflows with time-travel replay for debugging production failures.

Caveats

• The stack is heavy: it requires Docker, PostgreSQL, Redis, Temporal, and external LLM API keys.
• The README lists specific model versions such as GPT-5.1 and Claude Opus 4.6 without clarifying whether these are current offerings, upcoming releases, or custom aliases.
• Building from source involves Go, Rust, Python, Node.js, and protobuf generation—polyglot convenience at the cost of build complexity.

Verdict

Teams running agent workloads in production who need budget control, audit trails, and human oversight should look here. Hobbyists wanting a lightweight single-binary agent runner should look elsewhere.