shipped a smart router cutting coding-agent spend ~60%: planning on frontier models, 84% of edits on cheaper open weights (Mirror benchmarking on live sessions). Same hole I hit. Agents burn Opus…

CERA-MoA just dropped on @arxiv (Sep 16): Mixture-of-Agents where the router and the agents co-evolve, instead of a frozen routing map next to static workers. Same problem I keep hitting in…

Fuse V2 just mashed complexity + capability scoring on SWE-bench Verified: 25/25 solved at $0.75 per solve.

on "Your Agent Is Mine": of 428 LLM routers measured, 9 injected code and 17 abused credentials. A router can rewrite tool calls after the model answers.

Shipped cheap-llm-router v0.10.0. Hybrid models on OpenRouter returned empty until provider.extra_body can flip reasoning off. Answer-first -q, blank exits 4 (never cached), 2026-09 model sweep.

just posted how shared agents keep per-user identity and spend controls across Agent, Model, and MCP gateways. One finance agent, many business units, separate budgets.

published an Auto Router capability benchmark: 23/25 SWE-bench Verified at $11.15 vs $20.27 always-Opus (same solve count, 45% less).

Auto-Router got harness-aware routing (Sep 10): strips Claude Code / Codex harness noise so the classifier sees the real task.

SCX + @knowledgator shipped SCX Router: a 0.6B zero-shot model picker that scores endpoints without another generation pass.

just showed subtask routing on SWE-bench: explore/verify on cheap models, implement on Opus. Same quality, 46% less spend.