LangChain defines
- Prompts and chains
- Agent and graph structure
- Tool implementations
- Application state and user experience
// LangChain integration
Keep LangChain. Stop repeating proven work.
Keep the agent application your team already built. Punk finds repeated work, tests whether a proven result can safely serve it, and leaves uncertain work with the original model.
Punk is independent of LangChain. Product and framework names belong to their respective owners.
// complementary layers
Application logic stays application logic.
Punk contributes
- Compatible model traffic observation
- Identity- and policy-aware execution records
- Repeated-pattern detection
- Replay, shadow, reusable routes, fallback, and route explanations
// integrate in layers
Model calls first. Tools when declared.
-
Choose one chain or agent path
Use a workload with a stable owner, representative volume, and an output you can review.
-
Route supported model traffic
Follow the current adapter guide and confirm message, streaming, structured output, and error behavior.
-
Preserve identity
Attach tenant, app, agent, and pseudonymous subject context where relevant.
-
Declare tools and outcomes
Richer governance and evaluation require accurate action risk and success signals from the application.
Boundary: changing a model client does not automatically expose every graph transition, memory write, retriever call, tool side effect, or business outcome. Instrument those deliberately when the use case needs them.
Find the LangChain work your agents repeat.
Start with one measurable task. Punk will show what repeats, what could be reused, and what should stay with the model.