AI Signals Worth Watching: March 17, 2026
Three signals this week that are worth sitting with.
1. The agentic inflection is happening at infrastructure, not apps
The week of March 7–13 is already being marked as the moment “agentic AI” moved from technical vocabulary to operational reality. Not because of a single release — but because every major model dropped this month is designed around one question: can it execute multi-step tasks autonomously, or just answer questions?
GPT-5.4 shipped with a 1.05M token context window and a new “Tool Search” feature — the model dynamically looks up relevant tool definitions at inference time rather than loading everything into the prompt. That’s an architectural decision, not a feature. It’s designed for complex agent pipelines where hundreds of tools might exist but only a few are needed per task. NVIDIA’s Nemotron 3 Super, meanwhile, is the first open-weight model to hit 60.47% on SWE-Bench — the benchmark for autonomous software engineering — and it ships under an open license specifically so regulated enterprises can run it on-premises.
Both releases tell the same story: the infrastructure layer is maturing fast. Multi-step planning, error recovery, tool use, and autonomous execution are becoming table stakes, not differentiators.
The market data confirms the direction: agentic AI is projected to grow from $9.14 billion today to $139 billion by 2034. Gartner says 40% of enterprise apps will use agents by end of 2026, up from less than 5% last year. That’s not a gradual adoption curve — that’s a step function.
2. Google just locked in the default AI layer on 95% of smartphones
Apple’s reimagined Siri — launching with iOS 26.4 this month — runs on Google’s Gemini, not OpenAI’s models. The deal is worth $1–5 billion annually to Google and gives Gemini default access to more than two billion Apple devices.
Combined with Android, Google now owns the ambient AI layer on essentially every smartphone on earth.
This matters beyond the obvious. Google’s goal isn’t just to be the smartest model. It’s to become the default intelligence layer across every computing surface people actually touch — silently handling queries, interpreting context, surfacing information before users even ask. The Apple deal is a distribution play disguised as a technology partnership.
For OpenAI, losing default iPhone integration is a real strategic setback. ChatGPT stays available via opt-in, but defaults are where behavior lives. Sam Altman reportedly issued an internal “code red” when Gemini 3 launched. The OpenAI-Jony Ive hardware project — still unannounced — now reads differently: if you can’t win on existing devices, build the next device category.
Meta’s situation is more revealing. Their next flagship model, internally called “Avocado,” has been pushed from March to May because it underperforms competitors on reasoning, coding, and writing. They’re reportedly considering licensing Gemini as a stopgap. Even with Meta’s compute and talent, the frontier-model gap is widening. The top three — Google, OpenAI, Anthropic — are pulling away.
3. The second-brain space is filling up everywhere except where it matters
Two data points from this week. TwinMind — backed by Sequoia, founded by three ex-Google X scientists — has built an ambient AI that passively captures audio throughout your day and turns it into a personal knowledge graph. It processes on-device, stores only transcribed text, and works across 100+ languages. Strong privacy story, impressive technical execution.
At the same time, Tiago Forte’s Forte Labs just announced “The AI Second Brain” — a live cohort-based training program, enrollment opening March 26 — teaching people how to build a personal AI knowledge system manually. The fact that the productivity world is still packaging this as a skill course says something honest: the infrastructure isn’t self-serve yet.
Both products are solving for the same underlying problem — too much information, too many open loops, not enough retained context. But both stop at memory. Neither closes the loop. Neither acts. Neither knows that you said you’d follow up with someone, that a deadline is approaching, that two decisions you made last week now conflict.
The gap isn’t capture. The gap is coordination.
All three signals connect at the same point. The infrastructure race is producing better models, broader distribution, and more capable memory systems. What remains unbuilt is the layer that knows a specific human — their priorities, their relationships, their open commitments — and moves through their digital life with that context. Not a general assistant. Not a knowledge base. A coordination layer.
The window for building that is open. It won’t stay open long.
Eliran Keren — Founder of Deeplica, building the coordination layer for humans who’d rather direct than operate.
