Writing

Ideas, updates,
and hard-won lessons.

Engineering posts from the team building n00dles, plus every version we've ever shipped.

problemattemptn00dles
Featured
Why we built n00dles from scratch instead of patching LangChain
Six months, 4,000 lines of boilerplate, and one 2am incident later — we decided the existing options weren't good enough. Here's the full story.
R
Ramen Dass
·June 18, 2026·9 min
Engineering
Parallel agents with one line of code
Most LLM pipelines are embarrassingly sequential. Here's how n00dles' parallel() primitive works under the hood.
Soba Chen · May 316 min →
Production
State management in production LLM pipelines
Checkpointing, distributed locks, crash recovery, and why your in-memory state store will eventually betray you.
Udon Price · May 2211 min →
Opinion
The hidden cost of LangChain
It's not just the boilerplate. The real cost is the abstraction debt you accumulate every time you paper over a bad API.
Al Dente · May 87 min →
Tutorial
Building a deep research pipeline in 50 lines
Web scraping, parallel analysis, synthesis, and structured output — all wired with n00dles in one afternoon.
Fettuccine Romano · April 2914 min →
Deep dive
How we think about retries in multi-agent systems
Not all failures are equal. Transient errors, rate limits, semantic failures, and timeouts each need their own strategy.
Pad Thai-ler · April 148 min →
Release
Announcing n00dles v0.1.0 — public beta
After six months of internal development and two production deployments, we're open-sourcing n00dles.
Pho Nguyen · April 15 min →
v0.3.0
Jun 2026
Testing utilities — no network calls in tests
The last documented-but-unbuilt gap from the testing guide: agents could only be tested against a real provider or a hand-rolled fake.
  • NEWmock_agent() — stub one agent's LLM call with a fixed value, validated against its declared return type
  • NEWmock_pipeline() — stub an entire pipeline's execution at once, for tests that only care what calls it
  • NEWBoth work as decorators (@mock_agent(...)) as well as with blocks
  • IMPraises= goes through the agent's real retry/fallback handling instead of skipping it
v0.2.0
Jun 2026
Parallel execution and conditional routing
Fan-out and branching, the two biggest gaps from the v0.1.0 sequential-only release.
  • NEWparallel() and the | operator — concurrent fan-out with automatic result merging by agent name
  • NEWbranch() — route to one of several agents based on a classifier's output
  • NEWmax_concurrency on parallel() for rate-limit-conscious fan-out
  • IMPPartial-resume now checkpoints each member of a parallel() group independently
v0.1.0
Apr 2026
Public beta — open source release
First public release. Sequential composition only — see the post below for what was deliberately left out.
  • NEW@agent decorator with full type validation
  • NEWpipeline() and >> for sequential composition, plus run()/arun()
  • NEWProvider support for Anthropic, OpenAI, Mistral, Gemini, Ollama, and more via litellm
  • NEWRetry with exponential backoff + jitter, per-node timeouts, and fallback agents
  • NEWSQLite state store with checkpoint-and-resume, on by default
  • NEWPydantic v2 validation for structured agent outputs
  • NEWStructured trace events with an optional OpenTelemetry exporter
🔜 Circuit breaker, distributed (Redis/Postgres) state backends, Langfuse/Helicone exporters, and the noodles deploy CLI are all on the roadmap and described in the docs, but none have shipped yet — see the docs for current status on each.