John Nicev stared at his terminal last week, watching LangChain’s agent graph crumble for the umpteenth time — a human pause, and poof, back to square one.
Selectools. That’s the fix he built. And in a market bloated with AI agent frameworks promising the moon but delivering debugging hell, this one’s different. Lean. Pythonic. Production-ready from day one.
Look, the AI agent space exploded last year — OpenAI’s assistants API, Anthropic’s tool use, Gemini’s multimodal calls. Market dynamics scream opportunity: Gartner pegs agentic AI at $47 billion by 2028. But frameworks? They’re the bottleneck. LangChain dominates with 80k GitHub stars, yet pros whisper about its complexity. Nicev’s post lays it bare: every framework vows ‘connect LLM to tools and go,’ then reality bites.
Every AI agent framework makes the same promise: “connect your LLM to tools and go.” Then you start building.
He nails it. LangChain demands five packages for basics. LCEL’s | operator? Hides a Runnable protocol that nukes your debugger. LangSmith? Pay to trace your own code. LangGraph pauses? Restarts nodes. Brutal for production.
Why LangChain’s Production Promise Cracks
Nicev wasn’t tinkering — his team built agents for real customers, handling live requests. Existing tools? Demo darlings, not battle-tested. So selectools emerged: tool calling that just works across OpenAI, Anthropic, Gemini, Ollama. No adapters. No schemas.
Traces? Baked in, no SaaS. Every run() spills details: tools called, timings, costs. Guardrails ship standard — PII detection, injection blocks, topic filters. One config, everywhere.
Multi-agent? Plain Python routers, not DSLs or Pregel. A three-agent chain: result = AgentGraph.chain(planner, writer, reviewer).run(“Write a blog post”). Boom.
Composable pipelines: summarize | translate | format. Human-in-the-loop? Yields precisely, resumes without rerun. Deploy? selectools serve agent.yaml — HTTP, SSE, playground. Done.
Here’s my unique take: this echoes Requests vs. urllib2 back in 2012. Python HTTP was a nightmare — arcane, verbose. Kenneth Reitz’s Requests? Dead simple, exploded to ubiquity. Selectools does that for agents. LangChain’s the urllib2: powerful but punishing. In a world where devs crave simplicity (npm’s 2M packages, but pip’s lean winners thrive), selectools could claim 20% mindshare in lean-agent niches by 2025. Bold? Data backs it: 4,612 tests, 95% coverage, Python 3.9-3.13. Pre-launch audit fixed nine critical bugs via five-agent hunts.
Numbers don’t lie. 44 interactive docs with runnable examples. 40 real-API evals. 76 examples. 50 evaluators. 152 models with pricing. Apache-2.0. And a browser visual builder — drag-drop, export YAML/Python, zero install at https://selectools.dev/builder/.
Smaller than LangChain, sure. Young community. Need 50 integrations? Stick with the giant. But for staying out of your way — Python tracebacks, no paid logs — pip install selectools.
Is Selectools Production-Ready Right Now?
Short answer: yes, for most. Those stats scream rigor. Security audit? 56 findings, nine critical squashed. Stability badges in docs. But let’s poke holes — it’s niche. No massive ecosystem yet. If your agents juggle 100 tools, LangChain’s integrations win. Market share? LangChain’s at 70% in agent surveys (from LangSmith data leaks), selectools at zero-point-something.
Yet dynamics shift fast. Ollama’s local boom (1M downloads/month) favors cross-model plays like this. Anthropic’s Claude 3.5? Tools shine here, no gymnastics. Prediction: if Nicev nails community (GitHub https://github.com/johnnichev/selectools, docs https://selectools.dev), it’ll mirror FastAPI’s rise — from zero to Flask-killer in two years.
Corporate hype check: Nicev’s candid — admits LangChain’s safer for big shops. No spin. That’s refreshing in AI’s PR swamp.
But. Early adopters? Devs burning on LangChain will flock. Cookbook at https://selectools.dev/COOKBOOK/ has gems: agent pipelines, evals, deploys.
One line deploys crush Kubernetes rituals. In cost-conscious times (LLM tokens at $0.01/1k, scaling hurts), free traces matter.
Why Does Selectools Matter for AI Builders?
Agents aren’t toys. McKinsey says 45% enterprises pilot by 2024. Failures? Framework friction. Selectools strips it — Python functions for routing, | for pipes. Feels like shell scripting, but async, safe.
Parallel bugs fixed pre-launch? That’s enterprise DNA. Visual builder democratizes — non-coders drag nodes, export code. GitHub Pages host, free forever.
Skeptical? Run the numbers. LangChain’s bundle: 50+ deps, megabytes. Selectools: light, pip-friendly. Debugger heaven.
It works because it respects Python’s zen: simple beats complex.
🧬 Related Insights
- Read more: ExecuTorch Promises Voice AI on Every Gadget — But Does It Deliver for You?
- Read more: BunkerM Implants Local AI Brains into MQTT Brokers
Frequently Asked Questions
What is selectools?
Selectools is an open-source Python library for building production AI agents with simple tool calling, tracing, guardrails, and one-command deploys — no bloat, cross-model support.
How does selectools compare to LangChain?
Selectools is leaner, with plain Python orchestration, free traces, and no restarts on pauses; LangChain offers more integrations but heavier setup and paid debugging.
Is selectools free to use?
Yes, Apache-2.0 licensed, pip install, no subscriptions — even the visual builder runs free in-browser.
