Y Combinator’s Spring 2025 Request for Startups (RFS) identified Devtools for AI Agents as a major opportunity: software that helps teams build, test, deploy, observe, secure, and improve agents reliably. The RFS is no longer a live call for applications, but its thesis remains highly relevant in 2026 as companies move from agent demos to production workflows.
For Indian founders, the opportunity is not to build another generic chatbot wrapper. It is to solve the difficult engineering problems that appear when agents interact with business systems, sensitive data, multiple languages, and real customers.
What counts as devtools for AI agents?
An AI agent can interpret a goal, choose tools, maintain context, take actions, and recover from failure. Devtools are the infrastructure and developer-facing products that make those behaviours measurable and dependable.
A strong product may address one or more of these layers:
- Development: SDKs, agent runtimes, prompt and tool registries, workflow builders, and local testing environments.
- Evaluation: Test sets, simulation environments, regression testing, task-success scoring, and human review workflows.
- Observability: Traces showing prompts, tool calls, latency, token use, intermediate decisions, errors, and outcomes.
- Deployment: Hosting, model routing, queueing, state management, version control, rollbacks, and autoscaling.
- Reliability: Guardrails, retries, fallbacks, permission controls, schema validation, and deterministic components around probabilistic models.
- Security and governance: Secrets management, audit trails, data isolation, policy enforcement, and protection against prompt injection.
- Operations: Cost controls, capacity planning, incident response, and monitoring across providers and models.
The best products focus on a painful workflow for a defined buyer. “An all-in-one platform for agents” is usually too broad. “Regression testing for customer-support agents handling Indian languages” is easier to validate and sell.
Why the opportunity is still open in 2026
Large language models have become easier to access, but production agent systems remain difficult to operate. Model quality can change between versions; tool APIs fail; context windows grow expensive; and a successful answer in a demo does not guarantee a successful business transaction.
This creates demand for infrastructure that can answer practical questions:
- Did the agent complete the task, or merely produce a plausible response?
- Which tool call caused the failure?
- How much did each completed workflow cost?
- Can a new prompt or model be tested without disrupting customers?
- What information was accessed, changed, or exposed?
- Can the system work across English, Hindi, and other Indian languages?
Indian startups can build defensible products by combining technical infrastructure with local operating knowledge. Examples include evaluation for multilingual voice agents, compliance tooling for fintech workflows, and deployment systems optimised for cost-sensitive teams.
For a concrete view of production architecture, compare this category with building distributed systems with AI agents. The same concerns—state, retries, coordination, and failure recovery—shape serious agent platforms.
Product opportunities worth pursuing
1. Evaluation and testing infrastructure
Most teams still evaluate agents manually or with shallow benchmark scores. A useful evaluation product lets developers define business outcomes, replay real conversations, generate adversarial cases, compare versions, and require approval before release.
Start with one workflow, such as lead qualification, claims processing, or support resolution. Track task completion, escalation rate, factuality, policy violations, latency, and cost. Do not treat a high language-model score as proof of product value.
2. Agent observability and debugging
Logs are not enough when an agent makes several model calls and invokes external tools. Developers need a trace of the full run, including retrieved context, tool inputs and outputs, state transitions, retries, and final outcomes.
A differentiated product could automatically cluster failures, identify recurring tool errors, and connect technical traces to business metrics. Privacy controls are essential when traces contain customer conversations or financial information.
3. Secure tool execution
Agents need permissions, but many early systems give them excessive access. Infrastructure that provides scoped credentials, approval steps, sandboxed execution, and tamper-resistant audit logs can become a critical control plane.
This is especially relevant for Indian enterprises adopting agents in banking, healthcare, insurance, and government-facing processes. Security cannot be an afterthought added after the first enterprise pilot.
4. Deployment and cost management
Teams need to route requests between models, cache safely, manage queues, select regional infrastructure, and enforce budgets. A deployment layer can create value by improving reliability while reducing inference spend.
The product should expose unit economics clearly: cost per resolved ticket, completed onboarding, processed document, or successful voice interaction. Founders should also plan for model-provider portability rather than building an opaque dependency on one API.
5. Domain-specific agent infrastructure
Vertical tooling can be more defensible than generic orchestration. For example, a platform might provide evaluation datasets, connectors, policy templates, and observability for healthcare follow-up or restaurant ordering.
Products serving voice workflows should understand interruption handling, call transfers, speech recognition errors, and regional language variation. Resources such as multilingual voice agents for restaurants in India and patient follow-up with voice agents illustrate the operational detail a vertical platform must support.
What YC would likely look for
An RFS is a signal about a problem area, not a promise of funding. A credible application should show:
- A sharp customer and pain point: Name the team that experiences the problem and the costly workaround it uses today.
- A working product: A narrow, usable prototype is more persuasive than a broad architecture diagram.
- Evidence of demand: Pilots, usage, paid design partners, repeat workflows, or strong retention all help.
- Technical insight: Explain why the problem requires a dedicated product and why existing frameworks are insufficient.
- A credible founding team: Show direct access to users and the ability to ship infrastructure quickly.
- A large expansion path: Start with one wedge, then explain how it can become a broader platform or category leader.
Avoid presenting an agent framework as the company without explaining the buyer, distribution channel, and durable advantage. Open-source adoption can be valuable, but it must connect to a sustainable business model such as hosted infrastructure, enterprise controls, or premium observability.
A practical validation plan for Indian founders
Use a six-week process before investing heavily in platform scope:
1. Interview 15–20 teams already running agents or planning a production deployment.
2. Collect failure traces, not just feature requests. Ask for the last incident, manual workaround, and financial impact.
3. Choose one measurable workflow and define its success metric.
4. Build a narrow integration that works with the customer’s existing model and tools.
5. Run the product on real or carefully anonymised tasks, measuring reliability, cost, and time saved.
6. Ask for a paid pilot or a written commitment tied to a defined deployment milestone.
Teams that need rapid iteration can use rapid AI prototyping services for startups, but the prototype should test a real buying decision—not merely demonstrate that an agent can call an API.
India-specific design considerations
Build for the environment customers actually operate in:
- Support data residency, access controls, and audit requirements from the beginning.
- Treat multilingual input, code-switching, accents, and noisy audio as core test cases.
- Offer predictable pricing for startups and enterprises with strict budgets.
- Integrate with common Indian systems, including CRM, ticketing, payments, messaging, and identity tools.
- Provide deployment options that work across public cloud, private cloud, and controlled enterprise environments.
- Make human handoff and approval workflows first-class features.
For sensitive sectors, study the expectations behind HIPAA-compliant voice agents for hospitals, while adapting the controls to the relevant Indian legal, contractual, and industry requirements rather than copying a foreign compliance checklist.
Bottom line
The Spring 2025 RFS was directionally right: agent adoption creates a new infrastructure layer, and that layer is still being built. In 2026, the strongest opportunities sit where agents meet measurable business outcomes—testing, security, reliability, cost, and domain-specific operations.
Start with one painful workflow, instrument it end to end, and prove that your product improves a metric a buyer cares about. That is a stronger foundation for an accelerator application, enterprise sale, or grant proposal than a broad claim about transforming AI development.
FAQ
Is YC’s Spring 2025 RFS still open?
No. It was a Spring 2025 thematic request. Founders can still use its thesis to shape products and track future YC application cycles.
Do I need to build a new foundation model?
No. The opportunity is in the software layer around agents: evaluation, deployment, observability, security, orchestration, and specialised workflows.
What is the best first customer?
Choose a team already spending time on agent failures or manual review. A customer with a live workflow and clear operational pain is more useful than a broad market survey.
How should I differentiate from open-source frameworks?
Own a specific outcome, dataset, workflow, integration, or operational control. Framework compatibility can support distribution, but it is rarely sufficient as a moat on its own.
Explore AI funding and support
If you are building infrastructure for AI agents in India, combine product validation with a clear funding plan. Explore AI Grants India for relevant grants, programmes, and resources that can help turn an early prototype into a production-ready venture.