My 2026 agent stack after a year of rebuilding: models are easy, the layers around them break
Latitude_3420 · reddit · 2026-10-26
A developer shares the agent stack they rebuilt three times in a year, concluding the model is the easy part—the layers around it are where things break. Their setup: Claude for judgment/writing plus a cheaper model for sorting and extraction; Claude Code for long jobs, Claude Desktop for step-by-step steering; Agentik as a single MCP connection for search, scraping, keyword data and image models paid from one prepaid balance; plain markdown files for memory; a simple cron job for scheduling; and guardrails as API-key spending limits, not prompt rules—"the prompt is a request, the key is a limit." If starting over, they'd set guardrails first. Known weak spots: Agentik is hosted-only without custom provider keys, and the scheduling is held together with tape.
More from coding & agent
- A visual guide to the 7 layers of scalable system design — goyalshaliniuk · 2026-10-26
- Solo dev renders 52-shot cinematic game trailer with Claude Code on one laptop for a few dollars — freechoice · 2026-10-26
- Open-source Buddy app runs Codex CLI on Android, builds and installs a game from one sentence — 3amae404 · 2026-10-26
- Killed the vector DB for SQLite BM25 (0.35ms) and wrote a 3.6ns Rust sentinel for agent loops — OpenComplaint524 · 2026-10-26
- Official A2A CLI simplifies testing Agent2Agent servers from the terminal — bibryam · 2026-10-26
- Claude Code rebuilt its own ComfyUI pipeline to fix prop continuity in a 2-min AI film — Odd_Definition_9724 · 2026-10-26