Skip to content

Trends: April 2026

Headlines

Coding Agents Are the New IDE

Claude Code, Codex CLI, Copilot agent mode — all shipping production-grade agentic coding. The split: terminal-native (Claude Code, Codex, Aider) vs IDE-integrated (Cursor, Copilot, Windsurf). Terminal agents are more powerful; IDE agents are more discoverable.

MCP Goes Cross-Vendor

Anthropic's Model Context Protocol is being adopted beyond Claude. Cursor, Windsurf, Continue all support MCP servers. This means tools built for Claude Code work everywhere — a genuine open standard, not vendor lock-in.

Local Models Are Practical

Qwen 2.5 14B on Apple Silicon via Ollama: fast enough for real tasks (issue screening, code review triage, classification). Not for complex reasoning, but perfect for cost-sensitive batch operations.

Agent Frameworks: Still No Winner

LangGraph, CrewAI, smolagents, Pydantic AI, AutoGen — all viable, none dominant. The uncomfortable truth: most agent use cases don't need a framework. A 50-line tool-calling loop covers 80% of needs.

A2A Launches

Google's Agent-to-Agent protocol. Positioned as complementary to MCP: MCP = agents talk to tools, A2A = agents talk to agents. Early days — unclear if the ecosystem will adopt.

Notable Repos (Starred This Month)

Repo Stars What Why Notable
HKUDS/LightRAG 32.7K Graph-based RAG PostgreSQL/AGE backend, active daily merges
openai/codex ~5K Open-source coding agent OpenAI's answer to Claude Code, sandboxed execution
google/A2A ~10K Agent-to-Agent protocol Could reshape multi-agent communication

Papers Worth Reading

  • To be added as I find relevant ones this month

What I'm Watching

  1. MCP server ecosystem — how many third-party servers get built this quarter?
  2. Claude Code vs Codex CLI — feature parity race, which one developers actually use
  3. Local model quality — Qwen 3, Llama 4 releases expected soon
  4. RAG evolution — graph RAG becoming standard or staying niche?