Developer Tools: Agent Tooling

The production agent-framework toolkit has consolidated into named primitives: Pipecat (voice), browser-use (web navigation), Mem0 (memory), Composio (OAuth across 1,000+ apps), RAGFlow (document retrieval), Dify (visual workflow builder), with Mastra as the TypeScript-first option backed by 1.77M monthly npm downloads and YC. Human-AI collaboration tooling has expanded: Roughdraft.md provides a local-first Markdown review app for commenting and suggesting edits on files in a workflow designed around coding agents, and Nango (NangoHQ/nango, open source) enables building product integrations using AI capabilities with a framework that reduces manual API mapping. Security and research are now packaged as Claude Code / Codex plugins — the Codex Security plugin runs end-to-end appsec (threat modeling, finding discovery, false-positive validation, attack-path narratives), and Evo turns a codebase into a self-instrumenting autonomous-research loop. MCP has shifted from "a standard" to "a survival requirement": Linear's MCP server expanded from engineering into product management, and Tolaria ships an out-of-the-box MCP server so agents read/edit a Git-based plain-markdown vault with zero external integration. Free inference is mainstream — NVIDIA hosts ~80 models via free APIs that plug into OpenClaude, OpenCode, Zed, Hermes, and Cursor — and browser-tool selection is now a measured cost/perf line item rather than a default. Web-agent development is getting its own reusable distribution layer: Browserbase's catalog packages researched website playbooks as open skills, while Camofox attacks the bot-detection/token-cost layer. Agents are also getting their own identities (Sendblue's iMessage numbers) and their own operating systems: NovaStation's multi-lane personal AI OS and Hermes Agent v0.12.0's unified Kanban dashboard demonstrate the "AI-native command center" pattern where parallel agents claim and hand off tasks. The pace is high enough that best practices for coding agents on large-scale projects can invert within six months, making tool choice and operating protocol a moving target.