This week's signals
11 trends · updated weekly// the big one
Agent safety and data-privacy went front-page.
Copilot's MCP trust layer, Codex's wider rm detection, Claude Code's runaway caps, and the Grok Build repo-exfiltration scare all landed in the same two weeks. Safety is now a shipping feature, not an afterthought.
02Local inference keeps getting faster. Multi-Token Prediction speculative decoding merged into llama.cpp (1.4 to 2.2x speedups on consumer GPUs), Ollama v0.32.0 added an agentic "Chat, Code" mode, and local MoEs dominate r/LocalLLaMA.
03Verification is the bottleneck now. Code-gen speed is no longer the constraint; human and pipeline review capacity is. "Agents don't work for us" now means "our verification pipeline can't absorb the volume."
04Cost arbitrage graduated to core infra. Custom model routers are everywhere: Warp routers, Vercel AI Gateway, and Product Hunt winners. Cross-provider cost routing is now a funded category, not a hack.
05The agent governance layer is maturing. Copilot MDM and OTel, the Codex "writes" approval mode, and reviewable diffs in Devin all race to make autonomous agents auditable and enterprise-controllable.
06Model releases now reach app builders in hours, not months. GPT-5.6 and Grok 4.5 were usable in Cursor, Figma, and Codex the same week they shipped, propagation is near-instant.
07Price is the real headline. Terra at $2.50/$15, Grok 4.5 at $2/$6, GLM-5.2 at ~1/6 of GPT-5.5; frontier-tier cost dropped hard, so re-run your cost-per-task math.
08Agents are moving into the main client. Codex inside ChatGPT desktop, ChatGPT Work, and Replit's desktop app; the standalone agent app is collapsing into the tools devs already live in.
09Local + open frontier is real now. GLM-5.2 open weights plus Colibrì running a 744B model on a 25GB laptop mean self-hosting a frontier model without a GPU cluster is no longer a stunt.
10Physical AI enters the chat. Mistral Robostral Navigate and HF LeRobot v0.6.0 show model labs reaching into robotics.
11Token-efficiency is now a first-class metric. PACE, CompactionRL, MosaicKV and tools like code-review-graph's 82x reduction all target the same pain: agents burn too many tokens on context.