Where to Build in AI Agents: 25 Startups Solving the Hardest Problems in 2026 (7 Categories)
I don't usually write articles like this, but 2026 has been a pretty interesting year for AI agents.
Here are 25 startups building in this space that I'm particularly bullish on, grouped into 7 categories based on the problems they're trying to solve. I've arranged these categories in an order that makes it easier to understand how the space is developing and where things might go next.
Some of these companies are solving problems so boring that you wouldn’t even imagine they exist in the first place.
Things like giving agents proper business context, deciding what they're allowed to do, helping them operate existing software, handling payments, recovering from errors, and even insuring against the risks of autonomous agents.
For each category, I've looked at how the problem was handled before, what these startups are doing differently, how far they've come, and where I think there's still room for someone to build something better.
I've also shared a few ideas I'd personally consider building.
I hope you actually get something useful out of this instead of just bookmarking it and never opening it again :)
1. Let me know about your Business:
AI agents are still not able to work on personalized manner for companies even if you train them on companies Small Language Models they still didn’t achieve accuracy.
It’s important for these AI Agents to know about your business in actual sense, so that they can go through the raised tickets, Company Docs etc. and execute the commands on behalf of their user autonomously
The startups:
Zeroset’s product - Nebula, attempts to give AI agents an evolving representation of how an enterprise operates.
Instead of repeatedly searching disconnected tickets, messages and approvals, an agent can work with a structured representation of what happened, what changed and what’s still open.
Akshat Kannan and Will Zhang closed a $5.2M pre-seed this month. Memory benchmarks are where most of the context layer pitches won't satisfy the need , so I'm watching Zeroset closely.
Turnstone is building a local AI workspace where multiple agents share context from connected applications, email, calendars, documents and local files.
Simply putting: Slack for AI Agents
@aryanvij02 and @jaixbhatia met on their first day at Berkeley and bonded over college football before they ever wrote code together. They came through YC's Winter batch on the standard $500K. What I like about Turnstone is the decision to keep everything on the user's machine and run on whatever model you already pay for.
Markov solves a related but different problem: models need examples of how professionals actually operate software.
Its datasets capture screen recordings together with mouse movements, keyboard actions and annotations from real workflows.
Markov reports having sold more than 15,000 hours of computer-use data to frontier AI labs.
@DevvMandal was the youngest person on the team at Sarvam. @harishashok already built a robotics company and picked up an Emergent Ventures grant. Two founders out of Bengaluru selling the data the labs train on is the position I'd want to be in.
Operon is building a structured, queryable representation of industrial facilities so engineers and software agents can understand relationships between equipment and systems.
Chemical plants, refineries and energy facilities have enormous quantities of engineering diagrams, schematics and equipment documentation and that’s why Operon is focused on industrial knowledge.
Anderson Chen studied chemical engineering at NTU and is building Operon out of YC's Summer batch on the standard $500K.
Operon says it has already turned 10,000+ P&IDs into live plant graphs, which is a lot of unglamorous work to have behind you this early.
Where’s the Moat:
Markov produces training data to improve model capabilities.
Zeroset and Turnstone focus on giving deployed agents useful context.
Operon structures complex domain-specific knowledge.
The Moat yet to be discovered:
I’d explore organizational memory that can explain decisions, not just retrieve documents.
For example, an agent preparing a contract renewal should be able to reconstruct why a particular discount was approved, whether the approval is still valid and which source supports it.
2. Permission doesn’t permit to Act:
AI Agents hold many permits to call any tool do anything but whether they are doing it with right intention or not that’s the thing that matters the most.
How do you decide whether An AI Agent allowed to read and write Emails won’t compromise with your confidential emails, they have the same power to read both ordinary and confidential email.
The identity of the agent is not enough to determine whether a particular action should be permitted.
The identity of the agent is not enough to determine whether a particular action should be permitted.
The startups
Agentic Fabriq provides identities, scoped permissions, approvals and audit logs for AI agents.
Its approach is to place a governance layer between agents and the systems they access, rather than relying on shared API keys or separate authorization implementations for every integration.
@paulinazhxu and @matthew_xu23 met at MIT admit weekend, both left before their second year, and took YC's $500K in the Winter batch.
Decawork is building company-specific models that evaluate whether an agent’s proposed action is appropriate for its current task.
An AI Agent has right to delete files now deleting duplicate test records could be legitimate but deleting customer records to solve an unrelated support problem wouldn’t be.
@arxneo14 built an AI compliance platform at Barclays, and at his last startup he hand-wired credentials for every agent he shipped. @_sarthak4 built AI systems at NVIDIA that ended up running inside OpenAI and Meta. They raised the standard $500K in YC's Summer batch.
Traceforce approaches security through visibility and monitoring.
Its software observes AI activity across company devices, applications, models and tools. The company also introduced TraceGraph, which connects that activity into a queryable representation.
That helps security and IT teams investigate which agents are accessing what, identify risky behavior and understand AI usage.
Xia Hua and Varun Wadhwa worked together at Clumio, where Xia ran the ransomware-protection products and Varun built the distributed backup infrastructure underneath them. They took YC's $500K in the Summer batch. My conviction here is in the team even more than the category.
Where’s the Moat:
This isn’t an untouched market.
The emerging startup opportunity is often at the intersection of these systems: preserving user identity, enforcing task-specific permission and understanding a sequence of actions across several tools and that’s where the Moat is.
Where could another founder build?
One idea I’d investigate is a permission simulator for agent workflows.
3. Let me use the Software:
We have already developed plenty of software to interact with other software and help in our workflows ( Automation ) but these software standalone didn’t operate efficiently until you bring humans in the loop. We have started to sort this out with RPAs ( Robotic Process Automation ) but now few startups are taking it further by Making and Allowing AI agents to follow the workflow learn by itself and execute the smooth operation throughout the software whether it is to automate the repetitive interactions with software or learn the new user behavior across the software.
The startups
Zomma builds computer-use agents for financial-services operations.
It is starting with workflows such as cross-border transaction alerts, where analysts investigate potentially suspicious activity and document their findings.
Its agents operate existing software interfaces and gather evidence rather than requiring the customer to rebuild every integration.
@jihprobs is a second-time founder who was on the founding team of a $1B hedge fund and helped scale Asia's largest wealth-tech platform. Varun Arumugam came from verification infrastructure at Apple and Bronco AI. Cashel Fitzgerald built backend systems at 8am that handle $20B+ in yearly transactions. The three of them raised $500K in YC's Summer batch.
A team that has sat on the analyst side of the desk is the team I'd trust to automate it.
Trope focuses on enterprise resource planning software, or ERP systems.
Businesses use ERPs to manage finance, inventory, purchasing and operations. Installing one rarely means it’s ready to run every company process.
ERP implementations often require extensive configuration, testing, documentation and consultants.
Trope is building AI-assisted services and agents for this implementation and customization work.
@mattchowx and @victorvannara grew up in Vancouver and met at UBC. Matthew wrote safety-critical firmware for the Cybertruck, and Victor built the first AI support agent at Amazon Payments. They're in YC's Summer batch on the standard $500K, going after the $50B a year companies spend on consultants after they've already bought the ERP.
Taiga is automating medical billing workflows, including assigning billing codes, checking claims, submitting them to insurers and handling denials.
The old approach often involves practice employees, external medical-billing specialists and revenue-cycle management software.
Taiga is attempting to automate more of the process.
@nandaguntupalli and @AdamWax3 have been friends since third grade, and both grew up doing billing at their parents' medical practices. That is the most unfair founder-market fit I've come across this year. They raised the standard $500K in YC's Spring batch, and Taiga already says 98% of its claims go through without intervention, with a physician approving every code. Having spent years inside billing infrastructure, I respect anyone willing to take on medical billing.
Plena is developing an operating layer for specialty medical practices.
Its agents handle administrative work involving referrals, scheduling, document intake, procedural requirements and collections.
@eebadaeebada spent three years embedded in specialty clinics after Datadog and Shopify. Ahmed Al Mudarris took the MCAT, then chose AWS infrastructure over practising medicine. Both have a parent who is a doctor, and Plena's office sits inside Eyad's mom's clinic, where she was the first user. They came through YC's Spring batch with the standard $500K and have grown 17x in eight months to seven figures in contracted ARR, mostly through word of mouth. In healthcare that kind of referral growth is the hardest signal to fake.
Async builds specialized agents for small businesses, including law firms, clinics and real-estate businesses.
Its pitch is closer to hiring someone to perform a role than buying another software dashboard.
@imbensmith sold shoes and trading cards before he dropped out at 20 to build software. Connor Park has been building since he was seven. They raised the standard $500K in YC's Summer batch.
Async says it has completed 15,000+ tasks for customers like Clifton Law and Noah's Ark Early Learning Academy. Founders who learned to sell before they learned to code tend to build for the buyer, which is why I like this one.
Where’s the Moat?
Automation of repetitive and tedious task ( very nuance and subtle to find; actual pain points ) will be top priority for companies whose workflow revolves around traditional software tools and building in this space will eventually going to be profitable.
Where could another founder build?
I wouldn’t immediately build another general-purpose computer-use agent.
I’d look for one expensive workflow with a measurable outcome and a persistent exception problem.
For instance, an agent might automate invoice matching until a purchase order contains an unexpected price, a missing tax field or a partial delivery.
Today, a human often handles those exceptions manually.
A company that owns the exception-management layer across existing automation tools could have an interesting opportunity.
4. Let me Pay:
I've tried Instinct, FO, Muse and quite a few other personal AI assistants. One problem I kept running into, especially with their early versions, was payments.
I'd ask an agent to complete a task, it would go through most of the process, but the moment it reached a payment page, it would stop and ask me to take over.
Which makes sense. Most payment systems were built for humans to enter their card details, approve transactions and complete checkout. They weren't designed for an AI agent trying to make a purchase on someone's behalf.
This is slowly changing. A few startups are now building payment infrastructure specifically for AI agents, figuring out how agents can make purchases, get authorization and complete transactions without needing a human to step in every time.
The startups
Orthogonal helps agents discover APIs, call them and pay for individual requests.
@chrisspickett worked on payments at Coinbase and on usage-based billing at Vercel. @berasogut1 is a two-time ICPC World Finalist who worked on the reCAPTCHA and Maps APIs at Google. They met at McGill and raised a $4.3M seed in June, led by Franklin Bi at Pantera, with YC, Pioneer Fund and Decasonic in the round. Christian has already built usage-based billing for humans, and now he's building it for agents across 50+ vendors and fiat, x402 and MPP. I'm biased because this is our world at Flexprice, but this is the team I'd bet on to figure out per-request pricing for agents.
Maven enables conversational AI agents to collect payments.
For example, a customer speaking to a home-services voice agent could complete payment during the call instead of being transferred to a different system.
Maven handles payment collection and related PCI compliance requirements through its integration.
@brandondboehme (ex-Meta, Amazon) and Wasi Ahmed (ex-Amazon) both came out of Berkeley and went through YC's Winter batch, with Pear VC, Robinhood Ventures and Peak XV also backing them. Their launch line is the most accurate description of voice AI I've read: it can handle the conversation but can't complete the transaction. Maven ships a one-click Vapi integration, and Vapi bills on Flexprice, so I see from the inside why closing that gap matters.
Agentcard helps agents make online purchases.
It provides a payment-card vault and checkout-completion infrastructure, attempting to solve the problem of browsers getting blocked, sensitive card details being exposed and checkout flows breaking.
@keyserfaty ran Atlas, an expense card for global teams that Remote acquired, then became GM of Cards at Remote. She turned down Stripe to build Agentcard. @PipeAbellos is a second-time YC founder and was employee #1 at Rappi. They took the standard $500K in YC's Summer batch.
Where’s the Moat?
Google introduced its Agent Payments Protocol, AP2, in September 2025.
Stripe and Tempo introduced the Machine Payments Protocol, MPP, in March 2026.
So the opportunity is not simply that agents need payments. Established payments companies recognize that too.
The remaining problem is making payments work reliably across different agent platforms, merchants, payment methods and authorization requirements and that’s where the Moat exists.
Where could another founder build?
I’d investigate the financial operations that happen after an agent makes a payment.
Imagine an agent making dozens of small API purchases during one customer workflow.
Who keeps the receipt? Who knows which customer or task generated the expense? What happens if a network timeout causes the agent to retry the purchase? Who requests a refund when a tool fails?
A payment protocol may authorize and transmit a payment without solving every downstream accounting and operational problem.
One possible product would connect an agent’s intent, authorization, payment, receipt and final business outcome into one auditable record.
5. Let me Check:
AI Agents Completing the process isn’t enough. If a human get something done he is accountable for how efficiently the task was completed, what methodology he used but incase of AI Agents it become really challenging to navigate whether the task assigned to an agent was done in the manner as it was supposed to and if you trying to verify this you are only stretching your resource limits and there is no purpose of assigning that task to an AI Agent.
The startups
Rubric AI is building reasoning and verification infrastructure that uses expert judgment to help train and evaluate models and agents.
It focuses on areas where simply checking whether an output looks plausible isn’t sufficient.
The company describes expert-verified reasoning traces, task-specific environments and mechanisms for agents to verify intermediate steps or escalate to humans.
Pragya Saboo was a product lead at Oscar Health through its IPO, built computer vision for operating rooms at Apella and drove new product lines at Asana. She's building Rubric on YC's standard $500K from the Winter batch. Verification only works if the expert behind it is real, and Pragya has spent her career in places where being wrong costs more than a bad answer. That background is why I'm bullish on Rubric.
Mount is developing insurance for businesses deploying autonomous AI agents.
Its approach combines scanning and testing AI deployments, identifying operational risks and providing purpose-built liability coverage.
The premise is straightforward: companies may hesitate to give agents significant responsibility if they can’t assess or transfer the associated financial risk.
Fabian Amherd left ETH Zürich after building real-time object detection for industrial robots, and he had been running a web development company since 16. John Bachmann dropped out of high school and grew an AI-first media company to $400K in revenue at 18. They raised YC's standard $500K in the Spring batch. One of the eight risk categories Mount covers is runaway usage and infrastructure overage, a risk I spend most weeks thinking about from the billing side. Insurance is how a new risk becomes normal business, and these two got there early.
Risklytics focuses on insurance for emerging technologies, including robotics, physical AI and companies deploying autonomous systems.
Unlike Mount’s described insurance-carrier model, Risklytics describes itself as an insurance brokerage.
It helps businesses identify appropriate coverage, compare insurers and negotiate policy terms, including AI-related exclusions.
Samuel Gold and Alexander Risio are both on leave from Harvard. They joined YC's Summer batch with $500K to build a wildfire risk model, until a batchmate who launches satellites told them he couldn't get insured. They pivoted, got licensed within a week and now operate in 15 states.
Where’s the Moat?
AI Agents still not trust worthy enough to handle Financial Operations and whether they are assigned to execute one or not there’s always Multiple Humans in the Loop verifying the steps. Companies have one most important Moat here: Make the AI Agent intelligent enough to become accountable for the tasks assigned to them.
Where could another founder build?
I’d investigate an evidence and verification layer for high-stakes agent actions. The hard part is making that evidence trustworthy rather than merely collecting another set of AI-generated explanations.
6. Let me Buy:
I think software discovery is going to change quite a bit with AI agents.
Today, when I need a tool, I search for it, compare a few options, read the docs and decide what to use.
But as we start relying more on agents to get things done, they'll increasingly be the ones searching for software, comparing products and deciding which one to use.
And I think companies will have to start paying attention to this. Can an agent find your product, understand what it does and actually use it?
You could have a genuinely good product and still lose potential customers simply because an agent couldn't figure out how to work with it.
I can see this becoming a pretty important distribution problem for software companies.
The startups
Scope helps software companies measure and improve how AI agents discover and interact with their products.
It runs realistic workflows and observes when agents choose a company’s product, when they choose competitors and where setup or documentation creates problems.
@anandPa94 did interpretability research on closed-source models at Princeton before working on GEO and AEO as an ML engineer. He's building Scope out of YC's Spring batch on the standard $500K and already works with Blaxel and an unnamed decacorn. Our team spends a lot of time asking whether AI search can find Flexprice. Anand is asking the harder version: once the agent finds you, does it actually pick you?
Armature is also focused on agent-led software discovery, particularly developer tools.
It tests coding agents against realistic repositories to determine which tools they select and whether documentation, SDKs and integrations work correctly.
@Totzenberger is ex-Palantir and studied at Polytechnique. Louis Scremin led AI automation and MCP servers at Joko as an early employee. They're building out of Paris on YC's standard $500K from the Spring batch.
Where’s the Moat?
Until now, software companies have focused on getting discovered by people. They invested in SEO, GitHub, developer communities, marketplaces and documentation to make sure someone looking for a tool could find and try their product.
But now AI agents are starting to do some of that work themselves. They search for tools, read documentation, compare options and even try integrating them into workflows.
I think companies will eventually have to optimize for this too. Having good SEO or a popular GitHub repository might help an agent discover your product, but what happens when it tries to use it?
If your documentation is confusing, your SDK doesn't work properly or the agent gets stuck during authentication, it might simply move on to another tool.
There's an interesting opportunity here for startups helping companies understand how agents discover, evaluate and use their software.
Where could another founder build?
Rather than immediately building another AI visibility dashboard, I’d investigate whether companies can directly improve how well agents use their products.
For example, a testing tool could automatically identify broken SDK examples, missing CLI commands, confusing authentication flows and documentation that causes agents to select incorrect parameters.
7. Smooth Operator:
I think one of the biggest challenges with AI agents is making them efficient enough to actually use at scale.
An agent might be capable of completing a task, but if it's calling expensive models for every small step, taking too long to respond or repeatedly retrying failed actions, the whole thing becomes expensive very quickly.
And that's why optimizing individual model calls isn't enough. You have to look at the entire workflow, from choosing the right model for a task to reducing inference costs, latency and infrastructure overhead.
Some of the interesting companies I've come across in 2026 are working on exactly these bottlenecks. They're figuring out how to make inference faster, move data more efficiently and build better infrastructure for AI workloads.
I find this space particularly interesting because improvements here will directly affect how cheaply and reliably other companies can run their AI products.
The startups
Overshoot builds infrastructure for applications that process live video using vision-language models.
A model analyzing a static image has a different workload from one continuously observing a video feed and responding in real time.
Overshoot is optimizing the streaming and inference path for these applications.
@zakariaornot spent seven years building pricing algorithms at Uber before writing GPU kernels at Meta. His cousin Younes was a founding engineer at Cosmonio, which Intel acquired. The two grew up in Agadir and came through YC's Winter batch with $500K to build a real-time vision API that responds in under 200ms. Someone who priced rides at Uber's scale knows that cost per inference decides whether video agents ever leave the demo, and that is why I'm bullish on Overshoot.
Eridu emerged from stealth in March 2026 to develop networking infrastructure designed specifically for AI data centers.
Drew Perkins co-founded Lightera, which Ciena bought, and Infinera, which Nokia bought for $2.3B. He started Eridu with Omar Hassen, who led product at AppliedMicro and ran divisions at Broadcom and Marvell, and with Mike Capuano. In March they raised a $200M+ Series A led by Socratic Partners, John Doerr and Matter Venture Partners.
The bottleneck it’s tackling is straightforward.
A data center might have extremely powerful processors, but if information cannot move between them efficiently, some of that processing capacity is wasted.
Etched publicly unveiled its working inference chip and rack-scale system in June 2026.
@ubertigavin, Chris Zhu and Robert Wachen all left Harvard to build a chip. Gavin found the idea while writing microkernels at OctoML, and Robert, his college roommate, joined after the first two. Etched had raised $800M by June, when it showed working silicon, and in July it closed a $300M Series C at $10.3B led by Sequoia. The line that makes me bullish is $1B+ in signed contracts before the chip was public.
Humans& is taking a different approach.
Instead of primarily building infrastructure around existing models, it wants to develop AI that better supports collaboration and coordination between people and agents.
@ericzelikman wrote STaR and Quiet-STaR during his Stanford PhD, then worked on reasoning and agent RL for Grok at xAI. He started Humans& with his former advisor @noahdgoodman, Google's seventh employee @gharik and @YuchenHe07 from xAI. The $480M seed in January, at roughly $4.48B, was led by SV Angel and Georges himself
@ylecun left Meta to start AMI Labs with @lxbrun, who built Wit.ai, sold it to Meta and then ran Nabla. @laurentsolly, @sainingxie, @pascalefung and Michael Rabbat round out the founding team, and most of them came through Meta FAIR. They raised $1.03B in March at a $3.5B pre-money valuation, led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions. Yann has argued for world models over pure LLMs for years.
Its research focuses on world models that aim to understand and predict how the world behaves, rather than relying only on conventional language-model approaches.
Where’s the Moat:
This space is little complex to figure out it’s impact is directly over
other space because this single handly redefine the Infrastructure for AI.
GPU infrastructure, high-performance networking, specialized inference engines, model-serving platforms and AI optimization tools were already substantial markets but they still need operational reforms and that’s where the Moat lies.
After studying these 25 startups, there are 2 ideas I'd seriously consider building myself.
1. Exception handling for AI agents
Agents can automate workflows, but what happens when something goes wrong? I'd build a system that detects failures, figures out what needs human intervention and lets agents continue without starting over.
2. Financial operations for AI agents
As agents start making purchases and paying for APIs, someone needs to track where the money goes. I'd build infrastructure for tracking agent spending, handling duplicate payments, refunds and reconciliation.
I think both have a lot of room for new companies to build. Especially as businesses start trusting agents with more responsibility.