Skip to content

Roadmap

The MVP is feature-complete: a zero-dep SDK, a FastAPI + SQLite backend, a CLI, a React dashboard, live SSE streaming, ten rule-based loop detectors, universal OpenTelemetry ingestion, and a LangGraph adapter.

The full, always-current plan lives in ROADMAP.md. The highlights:

Recently shipped

  • Universal OpenTelemetry ingestion at POST /v1/traces — any OpenInference / OpenLLMetry instrumented framework, no LoopLens code.
  • LangGraph / LangChain adapter for tight in-process capture.
  • OpenAI Agents SDK adapter — a native TracingProcessor that captures handoffs and guardrails, not just LLM/tool calls.
  • CrewAI adapter — a BaseEventListener that captures crew delegation as handoffs, so a stuck crew trips handoff_bounce.
  • Handoff capture — transfer_to_<agent> tool calls become handoff_started events, so handoff_bounce fires through both paths.
  • Three more detectors — exact tool+args repeat, empty-result loop, and opt-in cost-budget alerts (LOOPLENS_COST_BUDGET).
  • Run comparison — a /compare view that diffs two runs side by side (health, metrics with coloured deltas, and each run's loop warnings).
  • PyPI release with a UI-bundled wheel, CI matrix (3.10–3.13), looplens doctor, and browser auto-open.

Next

  1. AutoGen and Pydantic AI adapters.
  2. Simple graph view — an agent/tool transition graph, after the timeline.
  3. Arbitrary node-to-node handoffs — beyond the transfer-tool convention, without flagging healthy loops.

Later

A graph view, semantic similarity via embeddings, replay-from-step, and an opt-in LLM-based diagnosis layered on top of the transparent rules.

Out of scope

LoopLens stays a loop debugger, not a general observability platform. No prompt-versioning suite, no eval-dataset product, no full OTel backend, no graph-first UI in the core.