Today’s lead · X
Google ships Gemini Decision 1 for closed-set multimodal scoring
Google is separating decision workloads from chat-style generation. Gemini Decision 1 is pitched for calibrated closed-set choices across multimodal inputs, with low latency and output tokens billed at zero. For builders, this is interesting anywhere classification, review, routing, or scoring has been awkwardly implemented with free-form LLM calls.

Top signals
10 moreTools & repos
4 selected
Nod by n8n
Nod is n8n’s open-source personal agent that runs on a private computer on your Mac or PC. It works inside apps and websites, asks before actions you did not request, and lets you choose the AI provider.
shadcn/improve
A small but timely repo idea: spend your best model on codebase audit and planning, then hand cheaper models the execution plan. That matches how many teams are trying to control coding-agent cost without losing quality.

CALL-E
CALL-E gives AI assistants a real phone line for US and Canadian calls. The confirm-before-dial flow is the right safety shape for booking, quotes, stock checks, and appointment changes that still require phone menus.
Skillbox
Skillbox is a shared library for agent skills across Claude Code, Cursor, and Codex. The useful angle is governance: follow GitHub sources, review diffs, scan versions for prompt injection, and hold risky updates before they reach machines.
Blogs worth your time
3 reads
OpenRouter’s practical guide to stopping runaway agents
A useful, numbers-backed playbook for agent guardrails: detect repeated tool calls, normalize loop keys, cap steps and cost, then add API-key credit limits. The examples show why step limits alone do not bound spend.

NVIDIA adds runtime AI inventory to GPU platform validation
NVIDIA’s AICR now includes k8s-aibom for Kubernetes AI bill-of-materials records. This is governance plumbing: track models, runtimes, containers, and evidence from supported AI workloads instead of relying only on intended cluster recipes.

OpenRouter prices the self-hosting break-even instead of hand-waving it
This post usefully reduces “should we self-host?” to load, token mix, provider pricing, and privacy constraints. The standout point: API pricing can beat rented GPUs unless traffic keeps hardware busy enough.
Community discussions
1 threadsBuilders ask what agents are actually useful for
The thread starts with agent fatigue, but the best replies converge on boring utility: parliament monitoring, meal planning, CSV cleanup, chores, and fantasy football. The lesson: durable value looks like narrow workflows, not elaborate multi-agent theater.
Funding & acquisitions
1 moves
Fractile raises $720M Series C for AI inference chips
Fractile’s $720M Series C is a huge bet on inference hardware, not just another model startup round. The company is selling speed for agentic workloads, with first chips and systems targeted for next year.
Bengaluru radar
1 events
Project Genesis
A free in-person PES RR Campus challenge using ChatGPT and Codex to design a civilization, build a webpage, and respond to a surprise crisis.






