Web Research Agent

LangGraphLangSmithLIVE

Decomposes a research question into sub-questions, fans out parallel web-search branches, then synthesizes a cited markdown report.

plan+synthesize → deepseek-v4-pro · research → deepseek-v4-flash
Live Timeline
Events will stream here when you start a run.
Idle
Portkey Gateway
LiteLLM → Portkey → provider · all model traffic routed here
LLM calls0
Cache hits0
Provider cost$0.00000000
Saved by cache$0.00000000
Cost precision: 8 dp — real per-call amounts are sub-cent. No rounding up.
Spotlight
Real measured runs · reviewer-approved · no invented numbers
LangGraphframework
Strengths
  • +Explicit StateGraph — the plan → research → synthesize flow is declared, not emergent.
  • +Send-based parallel fan-out — M2 traces show 3 research branches overlapping in time.
  • +Framework-portable model calls — every LLM call goes through LiteLLM → Portkey, no provider SDK imports.
Limitations
  • −One source per research branch — a single tool failure fails the whole run (no graceful degradation yet).
  • −Cache hits require byte-identical prompts — when a fetched page changes between runs, downstream calls honestly miss.
  • −Each run is a fresh subprocess — cold-start latency on every run.
LangSmithobservability
Strengths
  • +Run tree clearly visualizes parallel branches (verified in M2).
  • +Accepts vendor-neutral OTLP/OpenInference spans — no LangSmith SDK needed.
  • +Free cloud tier, zero local resource cost.
Limitations
  • −Cloud-only — traces leave your machine (scrubber strips sensitive payloads first).
  • −OTLP ingest maps some OpenInference attributes imperfectly vs its native SDK.
  • −Sharing run trees publicly requires per-project setup.

Want this architecture for your team?

Anguliyam AI helps engineering teams design and ship production-grade agentic systems — framework-portable, observable from day one, and built to run where your data lives. Let's talk about your use case.