Home
Jul 23, 2026
View All

Cursor Router, monday.com's Production Agents, and AI for Women's Health

·1 underrepresented voice

The Short Version#

Three things worth paying attention to today: Cursor shipped a model router that makes "auto mode" actually intelligent, monday.com published real production numbers on AI agent adoption (90% of engineers using coding tools monthly, 50%+ PR throughput increase), and Teresa Torres featured Hertility's AI-powered women's health diagnostic — a concrete example of AI closing a gap that traditional healthcare products left wide open.

Cursor — Cursor Router Ships in Auto Mode#

Source: https://cursor.com/changelog/router Credibility: High (first-party changelog entry, July 22, 2026)

What happened: Cursor shipped Cursor Router, an intelligent model routing layer that now powers Auto mode. Instead of you picking a model per task, the Router analyzes each request and dispatches to whichever model is best suited — frontier models for complex reasoning, price-efficient models for simpler work. The changelog describes two modes: Optimization mode (maximize output quality) and Savings mode (reduce credit burn). Users can toggle between them.

Key capabilities:

  • Automatic per-request model selection based on task complexity
  • Two configurable modes: Optimization and Savings
  • Transparent routing — you can see which model was selected per request
  • Powers existing Auto mode, so no workflow change required for current users

Why it matters for PMs: This is a meaningful UX shift. Right now, most AI coding tools ask users to choose a model, which creates decision fatigue and requires users to understand model tradeoffs they shouldn't have to care about. Cursor is betting that the right abstraction is "describe what you want" and let the system handle model selection. If the Router works well, it also compresses costs — the Savings mode signal suggests Cursor is thinking carefully about unit economics as usage scales. This is worth watching for anyone building AI features where model selection is currently a user-facing concern. The pattern is: abstract the complexity, expose the control.

Critical questions:

  • How transparent is the routing logic? Do users trust the Router's choices, or do they override it constantly?
  • Does the Savings mode materially degrade output quality on tasks that actually needed a frontier model?
  • How does this affect enterprise customers who have negotiated specific model access or have compliance requirements around which models touch their code?
  • Will Cursor publish routing data — what percentage of requests go to each tier — so PMs and teams can calibrate?

Action you could take today: If your team uses Cursor on a paid plan, toggle to Savings mode for a week and log any tasks where the output felt underpowered. That's your baseline for understanding where intelligent routing helps vs. where model choice still matters.

monday.com on AWS — Production AI Agents at Scale, With Real Numbers#

Source: https://aws.amazon.com/blogs/machine-learning/ai-teammates-how-monday-com-runs-production-ai-agents-on-amazon-bedrock/ Credibility: High (AWS ML blog, production case study from monday.com engineering team, July 22, 2026)

What happened: monday.com published a production case study on their "AI Teammates" program — autonomous agents built on Amazon Bedrock that handle real engineering workflows. The numbers in this post are unusual because they're specific and reported internally, not marketing claims: 9 in 10 "Builders" (their internal term for engineers) now use AI coding tools every month, up from roughly half a year ago. Per-engineer PR throughput is up more than 50%.

Key technical details:

  • Built on Amazon Bedrock, running at enterprise scale inside monday.com
  • AI Teammates are agentic — they handle full workflows, not just single tasks
  • Adoption jumped from ~50% to ~90% of engineers using AI coding tools monthly within a year
  • The post frames this as an infrastructure and culture story, not just a tooling story

Why it matters for PMs: Two things stand out here. First, the adoption trajectory: going from half to 90% of engineers using AI coding tools in roughly a year is faster than most enterprise software rollouts. That's not just product-market fit — it suggests a tipping point where social proof inside engineering orgs drives adoption faster than any formal rollout program. Second, the PR throughput metric (+50%) is the kind of number that changes internal conversations about AI investment. If you're trying to make the case for AI tooling at your company, this is the most credible benchmark I've seen in the data today. Third, the shift from "coding assistant" to "AI Teammate" that handles full workflows is a signal about where the category is heading — and what users will expect next.

Critical questions:

  • What's the distribution? If top 20% of engineers account for most of the productivity gain, the headline number flatters the median experience.
  • How did monday.com handle the change management side — what made adoption go from 50% to 90%? Tooling alone rarely does it.
  • What does "PR throughput" actually measure here — merged PRs, opened PRs, or something else? The metric is doing a lot of work in this post.
  • Are the AI Teammates handling high-stakes tasks (architecture decisions, security-adjacent code) or mostly routine work?

Action you could take today: Pull your own team's AI tool adoption data — what percentage of engineers used an AI coding tool in the last 30 days? If you don't have that number, that's the gap to close first before making any roadmap arguments about AI investment.

Teresa Torres / Hertility — AI for Women's Health: Bayesian Diagnosis and Scan Automation#

Source: https://www.producttalk.org/building-ai-for-womens-health-how-hertility-combined-bayesian-diagnosis-and-scan-automation/ Credibility: High (Teresa Torres's Product Talk, published July 23, 2026, featuring Hertility's product team)

What happened: Teresa Torres published a case study on how Hertility — a women's hormonal health company — built AI into their core diagnostic product. The approach combines Bayesian reasoning for diagnosis with automated scan interpretation. The framing is explicitly about closing a healthcare gap: women's health has historically been under-researched and under-served by traditional diagnostic tools. Hertility used AI to do what the healthcare system wasn't doing.

Key patterns:

  • Bayesian diagnosis: the system updates its probability estimates as new information comes in, rather than making a single upfront call — which maps well to how hormonal health actually works (symptoms shift over time)
  • Scan automation: reduces manual interpretation work and turnaround time, which is a core friction point in women's health diagnostics
  • The product team had to make explicit decisions about where to surface AI confidence levels vs. where to present outputs as definitive — a classic trust-and-transparency tradeoff
  • Discovery process involved working directly with clinicians and patients to understand where AI could actually close gaps vs. where it would create new risks

Why it matters for PMs: This is one of the cleaner examples I've seen of AI being used to address a gap that traditional products couldn't — not to optimize an existing workflow, but to create a capability that didn't previously exist at this price and accessibility point. Hertility couldn't have served their user segment without AI. That's a different strategic situation than "add AI to make our existing product better." The Bayesian approach is also worth understanding: for any health, finance, or other high-stakes domain where state evolves over time, a model that updates incrementally is often more appropriate than one that makes a single confident call. The question is whether users understand or trust that kind of probabilistic output.

Critical questions:

  • How does Hertility communicate uncertainty to users — and do users understand or trust probabilistic outputs in a health context?
  • Where did they draw the line between AI recommendation and clinical decision? That boundary is both a regulatory and a trust question.
  • What happens when the Bayesian model is confidently wrong? How does the product surface that possibility to the user?
  • Is the Bayesian approach a genuine product advantage, or does it add complexity that most users don't benefit from?

Action you could take today: If you're building in any domain where AI outputs have real stakes (health, finance, legal), map out where your product currently presents AI outputs as definitive vs. where it surfaces uncertainty. Then ask whether that choice was intentional or just the path of least resistance.

Quick Hits#

The Thread#

The abstraction layer is becoming the product. Cursor Router hides model selection. LangChain's Eval Skill automates evaluation design. monday.com's AI Teammates handle full workflows rather than task assists. The week's pattern is about moving the user-facing interface one level up — away from "which model, which prompt, which eval" and toward "describe what you want." The PM implication: if your AI feature still exposes model-level choices to users, you're probably a version behind.

Sit With This#

Cursor Router's two modes — Optimization and Savings — let users trade output quality for cost. monday.com's data shows 50%+ PR throughput gains at 90% engineer adoption, but we don't know if that productivity came from frontier models or efficient ones.

For your team: If you're building an AI feature that involves model selection (even implicitly), what's the right default — optimize for quality or for cost? And who should own that decision: the PM, the user, or the system?