When AI Stops Assisting and Starts Operating
Hey there, investors and innovators!
This week's crop of launches out of the current YC batch has a common thread: AI is being handed the parts of the job that used to require a trained human's judgment call, not just their typing speed. We're seeing autonomous systems take on the open-ended, iterative work of a research scientist; software step in to run and safeguard physical test infrastructure; agents shoulder the compliance-heavy back office of regulated finance; and products hand their own customization over to end users rather than a product team. Taken together, it's a reminder that this batch isn't just shipping smarter chat interfaces, but also pushing AI into the parts of a business that were previously considered too judgment-heavy, too regulated, or too physical to hand off.
What the Pattern Says
- AI is becoming the operator, not the assistant. Rather than sitting beside a human and suggesting the next step, this week's launches run the process end-to-end — designing and interpreting their own experiments, executing a defined procedure without deviation, or keeping a physical test rig running unattended. The common design choice is autonomy over suggestion.
- Auditability is turning into the actual product. As AI moves into regulated and safety-critical territory, "it worked" isn't enough — founders are building the paper trail as a first-class feature, not an afterthought. Every action, decision, and branch gets logged specifically so a compliance team or an engineer can reconstruct exactly what happened and why, which is quickly becoming a prerequisite for selling into finance, energy, and hardware.
- The product is starting to bend to the user, not the other way around. A parallel thread this week is software that lets the people (or agents) using it reshape the tool itself, rather than waiting on a product roadmap — a small but meaningful shift in who gets to decide what a piece of software does.
- AI's frontier is moving off the screen. Several of this week's companies aren't purely digital knowledge-work plays — they're extending into physical test benches and hardware environments, which fits a broader pattern we're watching across the batch: the easiest AI problems are largely solved, and founders are chasing the harder, messier, real-world ones next.
Eight Capital tracks every YC batch, cohort to cohort. Reach out if you'd like a deeper cut on any theme above.
Fund Manager Picks
rekursiv.ai
Autonomous AI scientists whose own breakthroughs accelerate the next
Why it's interesting
rekursiv.ai (YC Summer 2026) is building autonomous "AI scientists" — a system that proposes ideas, runs experiments, learns from the results, and repeats, with no human in the loop. In a few days of self-directed research it set a new Pareto frontier on ARC-AGI (75.5% on ARC-AGI-1, 17.5% on ARC-AGI-2) at 16×–11,600× lower cost than frontier LLMs at comparable accuracy, and became the first system to reach 100% Sudoku accuracy with a network trained only on input–output pairs. The thesis: AI progress is bottlenecked by ideas, not compute. Founders: Josh Dillon (13 years as a Staff Research Scientist at Google Research & DeepMind; creator of TensorFlow Probability, ~1M downloads/month; later led foundational-model pre-training at Luma AI) and Dan Kondratyuk (first author of VideoPoet, ICML 2024 Best Paper; co-developed the models behind Luma's Dream Machine).
Daqstra
AI-native orchestration for test infrastructure in physical R&D
Why it's interesting
Daqstra (YC Summer 2026) is building an AI-native orchestration layer for the test infrastructure behind physical R&D — aerospace, energy, and robotics. Today engineers rebuild test rigs and stitch together fragmented hardware (NI/LabVIEW, PLCs, custom DAQ, sensors) before every run; Daqstra unifies that hardware plus live and historical test data and exposes an API to orchestrate the next test. The company says an early user is the lead integration/test engineer for Blue Origin's New Glenn, alongside an energy-sector deployment. The founding team met testing liquid rocket engines in the Mojave Desert and has published research in the field. A durable hardware-integration moat with accumulating, proprietary test data — squarely in the "deep tech getting real" theme.
Rapidfolio
AI for fintechs and banks
Why it's interesting
Rapidfolio (YC Summer 2026) automates the back office for fintechs and banks — AI agents that run fraud investigations, KYB/KYC, loan underwriting, and the preparation and filing of Suspicious Activity Reports, with accuracy and a full audit trail. The wedge is the 10–25% of revenue that regulated financial institutions spend on compliance and operations headcount; the product is SOC 2 and built around deterministic, auditable outputs. Founded by Naveen Qureshi, a repeat YC founder who previously co-founded Sable (YC S19) — a banking startup for new-to-America immigrants that was acquired — now attacking the exact cost center he lived in.
Hubble
The intelligence layer for patient information
Why it's interesting
Hubble (YC Summer 2026) is building the integration and governance layer for healthcare AI — a single MCP server that connects AI agents to EHRs (Epic, Cerner, athenahealth), payers (UnitedHealthcare, Aetna), labs, and pharmacy networks, HIPAA-compliant, with direct APIs where they exist and browser retrieval where they don't. It's the plumbing every healthcare-AI company otherwise rebuilds from scratch — sold to digital-health builders, providers automating work, and health systems governing AI. The founding team includes a former tech lead at Amazon One Medical who led 50+ engineers on in-house EHR/AI tools serving 1M+ patients and built Amazon Care's data platform from zero to nationwide scale.
Vendo
Let your users build their own features on top of your product
Why it's interesting
Vendo (YC Summer 2026) is building an embeddable agentic layer that lets a software company's own end-users create and customize features on top of the product — describing what they want in natural language and having it built within the brand's guardrails, rather than waiting on the vendor's roadmap. The bet is that as agents make software cheap to generate, the winning products will be the ones that let customers shape the last mile themselves, turning a static app into a personalized surface for each user. It targets the perennial gap between what a product ships and what any individual customer actually needs. Founder(s) and early traction not yet independently confirmed.
