Real agent projects pushed one builder toward specialized agents, reranking and structured memory
AccurateCartoonist38 · reddit · 2026-07-21
After shipping a few agent projects, the author says the biggest lessons were architectural, not model choice.
Changes made after real-world use:
- Stop sending the full conversation to every agent; summarize and retrieve only relevant context.
- Split a single “do everything” agent into specialized agents.
- Prefer reranking over plain vector search for RAG.
- Store structured facts for long-term memory instead of every message.
- Add error handling and retries to tool calls.
- Separate frontend and agent backend for easier deployment.
They’re now experimenting with multi-agent workflows, long-term memory, Gmail/Calendar integrations, scheduled execution, and MCP versus traditional tool calling.
More from coding & agent
- Devin adds e2b sandboxes for remote agent execution — badphilosopher · 2026-07-22
- Hermes Agent Refactoring Proposal: Decoupling via Event Bus and Monorepo Slicing — Promptmethus · 2026-07-22
- ty now reads Pydantic config keywords and field metadata — charliermarsh · 2026-07-22
- Pensar Launches AI Security Agent to Autonomously Discover and Patch 0-Days — andriy_mulyar · 2026-07-22
- ty adds first-class Pydantic support, including strict and lax field handling — charliermarsh · 2026-07-22
- Google launches Gemini 3.5 Flash Cyber for CodeMender, with limited access for governments — GoogleAI · 2026-07-22