// the find
coze-dev/coze-loop
Next-generation AI Agent Optimization Platform: Cozeloop addresses challenges in AI agent development by providing full-lifecycle management capabilities from development, debugging, and evaluation to monitoring.
Coze Loop is ByteDance's open-source slice of its internal agent-ops platform: prompt playground and versioning, structured evaluation (eval sets, evaluators, experiments), and full-trace observability for LLM/agent calls, all in one Go backend built on their Eino and CloudWeGo (Kitex) stack. It's for teams running agents in production who want evaluation and tracing bundled together rather than stitching together separate tools for prompts, evals, and observability.
It actually covers the full loop in one deployable unit — prompt debugging, evaluator/experiment management, and trace ingestion with SDKs in three languages, rather than shipping just a prompt playground and calling it done. The infra underneath is real production machinery, not a toy: Kitex RPC services, RocketMQ, ClickHouse for trace storage, Redis, S3-compatible object storage — the kind of stack that's actually been load-tested at Coze's scale. Model access goes through Eino, so swapping OpenAI for Volcengine Ark or another provider is a config change, not a code change.
This is a ByteDance-internal stack transplanted outward — Kitex, hz-generated handlers, RocketMQ — which means standing it up yourself means operating several moving parts (MQ, ClickHouse, Redis, object storage) instead of one binary, a real cost for a team that just wants prompt evals. The README is upfront that this is a cut-down 'open-source edition' of a commercial product offering only 'core foundational feature modules,' with no doc spelling out exactly what's gated behind the paid version. The authors themselves warn in the README that public deployment carries real SSRF and privilege-escalation risk and tell you to assess that yourself — this isn't hardened for exposure out of the box. Deep docs live on the GitHub Wiki rather than in-repo, and the project is clearly Chinese-first (README.cn.md, CN-market model docs linked before the international ones), so English coverage may lag.