Building Useful Agents: SF Hackathon Shows the Power of Practical AI

In March, NEAR brought together over 100 developers in San Francisco—the AI capital of the world—for the inaugural Useful Agents Hackathon. Co-organized by NEAR AI, SF Compute, and Electric Capital, this 24-hour sprint challenged participants to move beyond theoretical AI capabilities and meme-driven viral trends by building agents that tackle real issues with quantifiable, repeatable results. This hackathon represented a crucial step toward NEAR’s vision of an ecosystem where blockchain and AI converge to create solutions built on practical utility and measurable impact.

From Capabilities to Measurable Performance

“If you can’t prove usefulness with metrics—success rate, speed, error reduction—no one will use it,” noted Anthropic’s Adam Wolff during the kickoff panel, alongside NEAR’s Illia Polosukhin, Electric Capital’s Avichal Garg, and SF Compute’s Evan Conrad. This perspective set the tone for the hackathon, moving the conversation from theoretical AI to concrete benchmarks and quantifiable performance indicators.

NEAR took a fundamentally practical approach in designing the event. All teams were required to provide quantitative evidence of their agents’ utility through benchmarks, error analyses, and reproducible testing environments. Instead of pursuing virality, participants concentrated on creating tools that resolve actual problems with demonstrable metrics that show lasting impact.

An Ecosystem of Innovation

The hackathon’s $20,000 prize pool attracted serious talent, with specialized bounties from partners including Coinbase CDP, Phala Network, xTrace, Nevermined, MIZU, Questflow, Pond, Silverstream AI, PublicAI, and Exabits. This collaboration created a robust environment where teams could leverage cutting-edge tech stacks to build solutions for specific vertical use cases.

Throughout the event, participants had access to workshops covering essential topics, such as NEAR AI Agent Hosting, Private Vector Databases, and Edge AI Data Agents. These sessions provided valuable technical guidance while fostering a collaborative atmosphere where ideas and solutions could flow freely.

Teams deployed their agents directly to the NEAR Agent Hub during the event, making them immediately available to users beyond the hackathon. Be sure to check them out on the NEAR AI Developer Hub.

Special thanks to our guest judges: Anna Turos, Manuel Del Verma (Silverstream), Lincoln Mur (Coinbase), Hang Yin (Phala), and Felix Meng (xTrace) for your time and insight.

Winners Circle: Solving Real Problems with Measurable Results

First Place: Postt ($5,000)

Zahidul Islam’s team created Postt, an AI-powered social media manager tailored specifically for startup founders and professionals focused on personal branding and thought leadership. The agent automatically generates engaging content, designs visuals, schedules publishing times for maximum reach, and analyzes follower engagement to optimize future posts.

The team’s benchmark results were impressive, showing a 92% approval rate by human evaluators, 35% average engagement increase, less than 2% error rate, and increased impressions by over 3,000% within one week on a test LinkedIn account.

When asked what he liked most about building with NEAR AI, Islam said, “I don’t have to worry about inference cost, and great support from the team.”

Second Place: Due Diligence Agent ($3,000)

Vijay Sithambaram and team developed DiligenceAI, a multi-agent system that performs comprehensive due diligence on potential investments. Their orchestrated system of seven specialized agents covers everything from initial screening to market analysis, competitor evaluation, team assessment, technical due diligence, report generation, and final decision recommendations.

The project’s benchmark results demonstrated that their solution is 85% faster than manual due diligence, identifies 92% of risks correctly, has less than 5% false positive rate, and analyzes 73% more sources than manual processes.

Third Place: NEAR Food ($2,000)

Nelson Lai created NEAR Food, an agent that finds free food events while effectively navigating anti-bot measures that typically prevent such automation. The system scrapes event listings, analyzes them to determine the likelihood of free food, and uses a two-agent approach to prevent anti-bot measures by mimicking human browsing behavior.

In performance testing, NEAR Food demonstrated 92% accuracy, correctly identifying 6/6 positive cases and 5/6 negative cases. The project earned additional recognition from Phala Network and xTrace for its novel approach to privacy preservation through Trusted Execution Environments (TEEs).

The Future of Useful Agents

What distinguished the Useful Agent Hackathon was its focus on benchmarks and measurable performance. But this is just the beginning of what’s possible with utility-focused AI agents.

Looking ahead, NEAR sees agents evolving from single-purpose tools to interconnected systems that seamlessly collaborate to solve complex problems. The multi-agent approaches demonstrated by several teams point toward a future where specialized agents combine their capabilities, similar to how teams of human experts work together.

Perhaps most intriguing is the potential for agent collectives to operate as autonomous businesses. DiligenceAI’s approach—with seven specialized agents working in concert to perform comprehensive due diligence—hints at a future where entire business functions could be handled by agent teams. We might soon see fully agent-operated businesses that can generate content, manage social media, perform market analysis, handle customer support, and even make investment decisions—all while continuously optimizing based on performance metrics.

As benchmarking frameworks become more sophisticated, we’ll move beyond basic metrics to holistic evaluations that consider not just speed and accuracy, but also resource efficiency, adaptability to new domains, and ability to learn from user feedback. The agents of tomorrow won’t just execute tasks—they’ll continuously improve their performance based on real-world usage.

Build Your Own Agent

Ready to create your own useful agent? The NEAR Agent Hub makes it easier than ever to develop, deploy, and share AI agents with measurable performance. Start building today at app.near.ai and join the growing community of developers creating agents that solve real problems with demonstrable results.

Whether you’re working on content creation, financial analysis, job applications, or something entirely new, the tools and infrastructure are ready for you to turn your ideas into deployed agents. What will you build?