Y Combinator's "Requests for Startups" for Summer 2026 highlights several key areas where AI is becoming foundational, driving innovation across various sectors, from agriculture and scientific discovery to enterprise operations and national defense.

One core area is AI for Low-Pesticide Agriculture, addressing the environmental, health, and economic unsustainability of chemical-dependent farming. The proposed methodology leverages AI vision for real-time identification of individual weeds and pests, coupled with cheap sensors and cameras for pervasive data collection. Precision robotics enable targeted plant treatment, reducing the need for blanket chemical application. Simultaneously, advancements in biology, such as microbes, peptides, RNA-based solutions, and engineered plants, offer alternatives to synthetic chemicals, allowing plants to self-defend and outcompete weeds. AI augments these scientific breakthroughs, enabling a shift towards higher yields and significantly reduced pesticide use, potentially by 90%.

Another focus is AI-Native Discovery Engines, aiming to accelerate the traditionally slow scientific discovery loop (hypothesize, experiment, interpret, repeat). The methodology shifts from AI as a research copilot to intelligent systems capable of running *closed discovery loops*. Frontier AI models, reaching PhD-level performance in scientific reasoning, assist researchers in proposing hypotheses, designing experiments, analyzing data, and suggesting next steps. In domains like drug discovery, materials science, and protein engineering, these systems iteratively propose candidate molecules, utilize automated labs for synthesis and testing, and feed results back to refine subsequent candidates, thereby automating and accelerating the entire design-make-test-analyze process.

The RFS also introduces AI-Native Service Companies, which fundamentally change the business model from selling software (SaaS) or AI copilots to directly selling the *service* itself, with AI performing the core work. This targets the vast service market, which significantly outweighs software spend, especially for outsourced services. Specific areas of interest include insurance brokerage, accounting, tax, audit, compliance, and healthcare administration. The methodology relies on increasingly capable AI models to execute complex tasks, replacing human-delivered services entirely rather than merely augmenting them.

AI Personalized Medicine envisions a revolution in healthcare delivery through intelligent agents. These agents analyze highly personalized health data—including diagnostic tests, genome scans, EHR data, and wearables information—to provide highly accurate, user-specific suggestions. This is enabled by the plummeting cost of personalized diagnostics (e.g., genome sequencing falling faster than Moore's Law) and the decreasing cost of N-of-1 genetic therapies like mRNA, along with increasing regulatory openness. The aim is to democratize access to personalized health assessments and treatments by leveraging abundant data and intelligence.

The concept of a Company Brain addresses the challenge of AI automation within enterprises. Current AI models are powerful, but their blocker is access to structured domain knowledge. The proposed "company brain" is a system designed to pull, structure, and keep current critical organizational knowledge scattered across disparate sources (human minds, emails, Slack, tickets, databases). This living map of how a company operates (e.g., refund processes, pricing exceptions, incident response) transforms raw data into an "executable skills file" for AI agents. This allows AI systems to perform tasks safely and consistently, acting as a crucial intermediate layer between fragmented company data and reliable AI automation.

Counter-Swarm Defense highlights an urgent need for advanced defense against rapidly evolving drone threats, particularly coordinated, cheap, and autonomous swarms. Current counter-drone measures are inadequate and costly, with the cost advantage heavily favoring attackers ($3M Patriot missile vs.$500 FPV drone). The methodology involves developing high-capacity interceptors capable of neutralizing many drones simultaneously, and software that fuses data from all sensors and defenders into a single real-time picture. Innovation in non-kinetic defenses (e.g., aerosols, streamers) and new attacks on the autonomy stack (as radio jamming becomes obsolete) are crucial. The winning solutions are envisioned as real-time distributed systems, more akin to Cloudflare than traditional defense contractors.

Dynamic Software Interfaces proposes a future where software interfaces are radically customized by users, moving beyond static, one-size-fits-all designs. This is enabled by coding agents becoming sophisticated enough for users to act as their own "forward deployed engineers." The methodology involves software companies shipping core "primitives" and design decisions, with the explicit intention that users, via their coding agents, will heavily modify the final interfaces to suit their specific workflows (e.g., an email client appearing as a task list for one user and an events calendar for another). This requires rethinking the entire software delivery stack, including how developers expose software to coding agents, whether through source code, machine-readable components, or programmable middleware.

Electronics in Space focuses on the need for increased compute capacity in space, driven by the expanding ability to launch payloads via reusable rockets. Specifically, there's a significant market for inference chips optimized for space environments—considering factors like mass, thermal management, and radiation hardness. These chips would process data onboard satellites, enabling more autonomous and data-rich space operations.

Hardware Supply Chain aims to dramatically accelerate hardware iteration speed in the US to rival Shenzhen's efficiency. The problem is not just supply chain gaps, but the weeks-long loop from design to physical part compared to Shenzhen's single-day turnaround. The methodology involves building startups that: produce parts significantly faster, enable rapid hardware iteration, and tightly integrate design, manufacturing, and logistics. This aims to create a dense supplier network and close coordination between design and production, allowing hardware teams to move orders of magnitude faster.

Industrial Capabilities in Space explores developing self-sustaining industrial operations beyond Earth, particularly on the Moon. The methodology involves extracting raw materials like silicon, aluminum, iron, and titanium through electrolysis. These materials would then be used for 3D printing complex structures from molten regolith, leveraging the Moon's low-gravity environment for more efficient fabrication without the need for support structures.

Inference Chips for Agent Workflows addresses the inefficiency of current GPUs for AI agent workloads. Unlike standard "prompt in, response out" inference, agents operate in loops involving tool calls, branching, backtracking, and maintaining context over many steps. This bursty, heterogeneous workload leads to low GPU utilization (30-40%). The methodology proposes purpose-built silicon designed specifically for the "agent loop," emphasizing fast context switching between models, native speculative decoding, and memory architectures optimized for KV caches that persist across an entire execution graph. The key lies in designing both the chip architecture and a compiler that can efficiently manage these complex agentic execution patterns.

SaaS Challengers seize the opportunity presented by AI's collapse of software production costs (10-100x reduction), which has eroded the moat of legacy SaaS incumbents. The methodology involves building AI-native alternatives to existing SaaS products. This can range from cloning products and selling them at a tenth of the price, to fundamentally rethinking workflows with AI-native designs (not just chatbots), bundling multiple point solutions, or creating open-source replacements monetized through services and hosting. The focus is on challenging large, seemingly invulnerable enterprise SaaS products like chip design software, ERPs, industrial control systems, and supply chain management, replacing them with modern, AI-first solutions.

Software for Agents anticipates a future where the "next trillion users" are AI agents, not people. The problem is that current internet software is designed for human interaction (visual interfaces, buttons), which is slow and brittle for agents. The methodology involves rebuilding major software categories explicitly for agents as first-class citizens. This means prioritizing machine-readable interfaces (APIs, MCPs, CLIs) over visual ones, along with thorough documentation to enable agents to programmatically discover, sign up for, and use tools without human intervention. The opportunity lies in building the foundational software stack that agents will depend on, rather than just building agents themselves.

Startups That Want to Sell to Huge Companies identifies a new opportunity for startups to directly engage with F100-sized enterprises, previously out of reach due to sales friction, product development time, and risk aversion. AI has changed this by: (1) making large enterprise buyers actively seek AI solutions; (2) enabling small teams to ship thoughtful, nuanced products quickly (months, not years) that are useful to large organizations; and (3) increasing awareness among enterprise leaders about the need to adapt with AI. The methodology encourages founders to target these "most famous, important customers," as multi-million dollar deals and pilots are now feasible within a startup's first year.

Finally, Supply Chain 2.0 for Semiconductors addresses the extreme complexity and lack of visibility in the semiconductor supply chain, where a single advanced chip involves ~1,400 process steps across a dozen countries over five months, currently managed with outdated tools like spreadsheets and phone calls. This leads to bottlenecks (e.g., TSMC's advanced packaging, HBM memory shortages), and makes real-time allocation, multi-tier risk monitoring, and export compliance nearly impossible. The methodology calls for startups to build next-generation tooling that provides real-time allocation tracking, multi-tier risk monitoring, and export compliance, requiring a deep understanding of wafer allocation and packaging constraints. This is particularly crucial with new fabs being built in the US, demanding new supply chain infrastructure.

The AI Operating System for Companies identifies that the most effective AI-native companies make their entire operations queryable, turning an "open loop" into a "closed loop." This means recording meetings, tracking tickets, capturing customer interactions, and making all data legible to an intelligence layer that learns and adjusts. Currently, achieving this requires complex integration of disparate tools (Slack, Linear, GitHub, Notion, etc.) with custom code. The opportunity is to build a "connective layer" product that natively integrates these artifacts into a single intelligence layer, allowing AI to reason across company data, identify issues (e.g., engineering building the wrong thing), and generate specs for agents to execute, effectively creating a self-improving operational loop.