Stacks/Multi-agent orchestration

A policy engine plus unified traces across LangGraph and CrewAI (HN)

ps3544's team runs LangGraph and CrewAI side by side in production and solved two operational problems in-process: a deterministic Python policy engine that can block tool calls before they run, and a unified @trace decorator that writes every framework's spans into one SQLite database so cross-agent races can be queried like any other trace.

Use case

A multi-player production multi-agent system that mixes frameworks (LangGraph and CrewAI) and enforces what agents can and cannot do before tool calls run

Category

Multi-agent orchestration

Team size

not stated

Orchestrator / framework

Policies are enforced with an in-process pre-execution hook: before a critical action such as a tool call, the agent's intent is passed to a lightweight policy engine of deterministic Python functions (user role, budget, and so on). A BLOCK result raises an exception and the action never runs. The author compares it to a sub-millisecond, sidecar-free, framework-agnostic OPA. The same team runs both LangGraph and CrewAI agents.

Memory

Shared-data issues across agents and frameworks are treated as an observability problem rather than a separate memory product. Details of the data store itself are not given.

Evals

not stated

Observability

A unified tracing system with a @trace decorator that works across frameworks. All traces are written to a single local SQLite database, which the team queries (for example, all spans for a data_id ordered by time) to debug race conditions and interoperability failures. The approach is open-sourced in early beta as a toolkit called Clearstone, with the repo linked from the author's HN profile.

Monthly cost

not stated

Models

not stated

Observability tools

Custom SQLite tracer (Clearstone)

Other tools

In-process policy engineClearstone (open source, early beta)

Source

https://news.ycombinator.com/item?id=45733665

ps3544 · Oct 27, 2025