The promise of AI agents seamlessly connecting to APIs and handling complex business tasks autonomously sounds compelling. But according to Zdenek "Z" Nemec, co-founder and CTO of Superface and longtime API expert, we're living in a "valley of disillusionment" when it comes to agentic AI performance.
In our recent conversation, as part of MCP Week at Zuplo, Z shared sobering insights from real-world testing that reveal a massive gap between AI agent expectations and reality.
If you'd prefer to watch Martyn & Z's conversation, you can in the video below!
The Harsh Reality of Agent Performance#
Superface's recent benchmarks show that even simple CRM tasks, like creating leads in Salesforce or updating pipelines in HubSpot, fail up to 75% of the time when agents attempt them repeatedly.
When testing six basic sales tasks across multiple runs, success rates plummeted dramatically. While a single execution might succeed 50-60% of the time, running the same task set repeatedly dropped success rates to as low as 10-20%.
This reliability problem isn't just a minor inconvenience, it's a fundamental barrier to deploying agents in production environments.
So, what can you do to try and achieve greater success here?
Building Better AI to API Connections#
Start with Narrow, Specialist Agents#
Rather than building one super-agent that handles everything, a microservices approach to AI agents delivers significantly higher success rates. "Specialist agents" that focus on specific domains or tasks can be optimized for particular business processes and API patterns, reducing the complexity burden on any single agent.
Limit Your Tool Count Strategically#
The optimal range for reliable agent performance is 10-20 tools maximum. Exposing hundreds of API endpoints as tools overwhelms current LLMs and destroys success rates. Focus on the core API calls needed to complete specific use cases rather than comprehensive API coverage.
Context and Planning Matter More Than You Think#
Simple requests like "book me a meeting when I'm available" require agents to understand time zones, working hours, and calendar contexts before making the actual booking API call. Most failures happen because agents skip these prerequisite steps or forget them in subsequent runs.
API Documentation Format Is Less Important Than Content#
Modern AI systems can work with pretty much any documentation format, OpenAPI, Markdown, or plain HTML, as long as the essential information is present. What matters is documenting business logic, authentication schemes, endpoint relationships, and the specific sequence of calls needed for complex operations.
Design APIs with Agent Consumption in Mind#
APIs optimized for agent use need careful consideration of response sizes and data selection. Features like GraphQL's selective field querying become crucial when dealing with context window limitations and token costs.
Authentication and Real-World Complexity Aren't Solved#
While new advancements like MCP provide a transport layer for connecting agents to APIs, it doesn't address fundamental challenges like authentication flows, rate limiting, error handling, or the complex business rules that govern real API usage.
That still lands in the hands of developers. Fortunately, with Zuplo's Model Context Protocol support, ensuring the endpoints you expose as tools are secure, rate-limited and erroring correctly comes as standard.
The Path Forward#
Technology isn't magic, and simply wrapping APIs in MCP servers won't solve reliability problems. Success requires thoughtful design at every layer, from model training and prompting to API design and tool description optimization.
The companies that will succeed in the agentic AI space are those that acknowledge this reality gap and systematically address the engineering challenges that make agents reliable enough for production use.
The future of AI agents isn't about building something that works once, it's about building something that works consistently, every single time, hundreds of thousands of times per day.
By focusing on reliability, narrow specialization, and careful tool design, we can build agents that actually deliver value in real business scenarios.
Many thanks to Z for taking the time to talk to me about this. For more details on their suite of agentic tools, and further reasearch in this space, head to the Superface website.
Have thoughts on this topic? Want to talk to us about our new remote MCP Server support in Zuplo? Reach out to in the #mcp channel of our Discord. We'd love to hear from you!