Which framework adapter should ship next?

[fitz2882]

Six adapters exist today — LangGraph, CrewAI, AutoGen v0.4+, LangChain, OpenAI Agents SDK, and the Claude Agent SDK. Each is 100–200 lines and does the same thing: drive the framework's iteration, call your error_fn on each event, stamp framework= on telemetry.

What's missing? Reply with the framework and, more usefully, what your loop looks like in it — where the iteration boundary is and what you'd count as the error. The iteration boundary is the part that varies and the part that takes the time to get right.

On my list, unranked: Pydantic AI, DSPy, LlamaIndex Workflows, smolagents, Agno, Semantic Kernel, Mastra, Haystack. Upvote the ones you'd use, or name one that isn't there.

If you don't want to wait: the raw observe() API works against any iterable, and the adapters are thin enough to crib. Copy whichever of loopgain/integrations/{langgraph,crewai,autogen,langchain,openai_agents,claude_agent_sdk}.py is closest in shape to your framework.

[LortuArte]

Whichever framework adapter you decide to ship next, keep an eye on execution latency and budget containment for agentic tool calls.

Most of these frameworks (LangGraph, CrewAI, AutoGen) suffer from a critical blind spot in production: if an agent hits a fast async loop, relying on cloud gateways to cap the spend is too slow. The network latency (~150ms) guarantees the budget drains before the external API block triggers.

We recently open-sourced AEGIS (pip install aegis-core-lortuarte-sdk) to solve this. It’s a pure Python L3 Policy Gate that applies ACID locks locally in memory (< 1ms).

Might be worth considering an adapter for local execution locks, or at least keeping the architecture in mind as you build out the ecosystem integrations!