Reverse-engineered from what YC (Fall/Summer 2026 RFS) and Accel are funding — because that's what the companies hiring you are building.
Why this list looks different: models are commoditised. The moat moved to harnesses, evals, agent-native infrastructure, and physical-world data. Hiring managers in 2026 scan for production signals — how you handle failure, what you measured, what you deleted. A repo with a published eval score beats a demo video.
Instant reject pile: wrapper chatbots, chat-with-your-PDF, to-do app with a GPT button, sentiment analysis on tweets.
Weekend to two weeks. Each one teaches a 2026-relevant primitive, not a tutorial.
Expose a niche API (your gym's booking system, a transit feed, a university portal) over Model Context Protocol. MCP went to the Linux Foundation in Dec 2025 and is now the default integration layer.
→ Signal: you can design an API for agents, not humans. Almost every AI team is doing this right now.
A CLI that runs prompts against a fixed test set on every commit and fails the build on regression.
→ Signal: the single biggest gap in the industry — 89% of teams with production agents have observability, only 52% have evals. Say that number in an interview.
Liveness challenge + device attestation + a signed proof, drop-in. A finance worker recently wired $25M to a video call where every other participant was a deepfake.
→ Signal: security thinking + cryptographic primitives. Rare in junior portfolios.
Watches an API's OpenAPI spec, detects breaking changes, posts a plain-English summary to Slack.
→ Signal: structured extraction from noisy text, and that you understand what actually breaks production.
Not an assistant. One job — logging medication, capturing site notes, stock counts — on-device.
→ Signal: latency budgets and offline inference. Everyone can call an API; few can hit a latency target.
Put a cheap sensor somewhere nobody measures (air quality on your street, soil moisture, noise) and publish a clean, versioned, public dataset with an API.
→ Signal: real data engineering, not Kaggle. And you now own data nobody else has.
Take a site built for human eyes, generate the machine-native surface: structured endpoints, llms.txt, a CLI. YC's line: "Machines do not require GUIs."
→ Signal: you understand why browser-clicking agents are a hack.
Auth, DB, domain, done. YC's A Cloud for Small Software RFS: small software is now trivial to build and still painful to ship.
→ Signal: DevOps and the boring 80% of software. Interviewers probe this and most candidates fold.
Two to six weeks. This is the tier that actually converts to offers — big enough to have hard tradeoffs, small enough to finish.
Multiple people watch, redirect and hand off the same long-running agent, Figma-style. YC's Multiplayer AI RFS: AI hasn't had its Google Docs moment.
→ Signal: realtime systems (CRDTs, streaming) + product taste. A very strong senior-track project.
HVAC, tree surgery, marine repair — quoting, scheduling, invoicing, routing work between a human and an agent. YC funded Kebra (S26) for exactly this.
→ Signal: full-stack shipping against a real workflow. Best project if you're targeting vertical-AI startups (there are hundreds hiring).
Watch a regulator's filings, diff them, map changes to obligations, generate the audit trail. YC's AI-Native Compliance Infrastructure.
→ Signal: you can build something auditable and explainable, not vibes-based. Fintech and regtech hire on this.
Vendor ships a breaking change → agent scans the codebase, finds affected usages, opens the PR. From YC's Self-Maintaining APIs (30%+ of AWS service downtime traced to unnoticed external changes).
→ Signal: AST parsing, codegen, safe automated changes. Dev-tools companies will interview you on sight.
Most "memory" is a vector store that only grows. Build decay, contradiction resolution, and provenance on every fact.
→ Signal: context engineering — which has replaced prompt engineering as the core AI-engineer skill.
Not flashcards. Model what the learner misunderstands, generate the next task from that. YC's The Primer RFS.
→ Signal: you can define and measure a success metric (learning gain) instead of shipping a feature.
Mixed fleet of agents, robots and people with wearables — decide who does what, and prove the routing beats a naive baseline.
→ Signal: optimisation + simulation. Strong for robotics and ops-heavy companies.
Escrow, spend limits, revocable delegated authority. YC calls agentic commerce inevitable now regulatory clarity landed. Testnet only.
→ Signal: you can reason about trust boundaries when the actor isn't a person.
Trace an agent run, attribute tokens/latency/dollars per step, flag where a smaller model would do.
→ Signal: you think about unit economics. Instantly separates you from candidates who've never seen a bill.
Assimilate your sensor network into a short-horizon forecast and score it against the official one. Sorcerer (YC) does this with weather balloons; the US government buys the output.
→ Signal: honest backtesting and a real ML pipeline end to end.
Months. One of these done well is a standout thesis, a research role, or a company.
Any agent run replayable exactly — same tool results, same branch points — for debugging and regression.
→ Signal: distributed systems, event sourcing, sandboxing. This is the missing layer in production AI. Infra teams will fight over you.
Hypothesise → design experiment → run → interpret → update, in one narrow domain where experiments are cheap. rekursiv.ai (YC S26) is building this.
→ Signal: RL and credit assignment. Research-engineer track.
High-fidelity simulator + reward spec a frontier lab could train against. Maingen (YC S26) does this for factories. Value is in fidelity, not the RL.
→ Signal: domain modelling and sim-to-real. Frontier labs are actively hiring for exactly this.
Agent traffic is bursty and branchy — nothing like chat. Speculative branch execution, KV-cache reuse across sibling agents, preemption.
→ Signal: memory management and systems performance. Highest-paid niche in the industry.
UI that assembles itself per user and task at runtime. YC's Dynamic Software Interfaces. Hard part isn't generation — it's stability, accessibility, and not confusing the user every session.
→ Signal: program synthesis + HCI rigour. Very few people can do both.
Medicine, law or finance: every claim traceable to a source, unsupported claims structurally blocked. Measure hallucination rate honestly and publish it.
→ Signal: evaluation rigour and red-teaming. Publishing your own failure rate is the strongest credibility move available to you.
Perception + decision stack with no cloud, degraded GPS, tight power. The constraint set behind YC's defense RFS and Accel's defense-tech allocation. Build the civilian version: search and rescue, agriculture, disaster response.
→ Signal: embedded systems, sensor fusion, model compression. Defense and robotics money is enormous in 2026.
Pattern across all 25: each produces proprietary data, a hard systems artifact, or a real workflow — the things models can't commoditise, and therefore the things companies still need to hire humans for.
If you want to see prompts for shipping some of these fast, there is a single-file sticky-notes app, a frosted-glass refraction cube in raw WebGL, and a drivable 3D car game.