The development era shifted from chat UIs to agentic systems that call tools, read repos, and execute workflows. Here's the stack I actually ship with in 2026.
LLM landscape (what I use and why)
| Model family | Best for | Production notes |
|---|---|---|
| GPT-4o / o-series | General reasoning, multimodal | Strong tool calling; watch token cost at scale |
| Claude (Anthropic) | Long context, careful codegen | Excellent for refactors and spec adherence |
| Gemini | Google ecosystem integrations | Good when GCP-native |
| Llama / Qwen (via Ollama) | Local-first, privacy, job automation | Zero cloud bill; tune prompts per hardware |
| Open-weight (HF) | Fine-tune & domain agents | Pair with LangGraph for guardrails |
No single winner — I route by latency, cost, privacy, and tool compatibility.
Model Context Protocol (MCP)
MCP standardizes how agents connect to databases, APIs, and IDEs. Instead of bespoke integrations per tool, you expose resources + tools once. This is why Cursor, Claude Desktop, and platform teams are converging on MCP servers for:
- Postgres / commerce APIs
- GitHub PR workflows
- Internal admin panels
Agent patterns that survive production
- Graphs over chains — LangGraph state machines with explicit human gates
- Tool allowlists — never give agents open network access
- Observability — log every tool call with correlation IDs
- Fallback paths — keyword routers when LLM confidence is low (social commerce order lookups do this well)
AEO / GEO / llms.txt
Discovery is no longer только Google. Answer engines (Perplexity, ChatGPT browse) and llms.txt files help models cite you accurately. I ship structured data + machine-readable profiles on every client site now.
Takeaway
Treat LLMs as orchestration layers, not magic. The engineers who win in 2026 pair models with MCP, graphs, and boring reliability engineering.
