Interrupt day 2: Agents in action

The second day of the Interrupt conference shifted focus from technical fundamentals to real-world implementations. The presentations revealed a clear pattern: the sectors leading AI agent adoption are those that are data-intensive and have large user bases. Banks, telecommunications companies, logistics firms, and software platforms dominated the presented use cases, showcasing implementations that handle massive interactions or highly complex processes.

Meanwhile, the company and startup booths offered a fascinating complementary perspective on the ecosystem emerging around agent development. I was able to speak directly with creators of specialized tools: IDEs designed specifically for agent development, authentication and digital security services adapted for agents, platforms for API connections and MCP protocols, experimental generative interfaces, vector databases optimized for agent memory, and even deployed agent graphs that allow understanding of their logic and internal functioning. These conversations revealed a highly specialized tool ecosystem that is rapidly maturing to support the development and deployment of agents in production.

During the presentations, I noticed the absence of use cases in academia, scientific research, public sector, or environmental applications—areas where agents could have transformative impact. The trend continues: sectors with greater capital and commercial urgency continue to lead the development of cutting-edge systems, while areas of potentially greater social impact lag behind in adopting these technologies. However, the presented implementations offer valuable lessons on how to translate technical concepts into systems that actually work at massive scale. Below, I analyze the talks that best represent these emerging patterns and the fundamental challenges that agents face in production.

Experience and Precision
‍
The presentations by Carlos Pereira from Cisco Systems and Zheng Xue from JP Morgan Chase perfectly illustrated how different industries approach agent implementation, but also revealed common patterns that go beyond the specific sector. Carlos described how Cisco uses agents to create customer experiences that he qualifies as "unimaginable" for traditional companies, while Zheng focused on agent systems for real-time investment decisions where millions of dollars are at stake.

What's fascinating about both approaches is the architectural sophistication they require to function in production. Cisco has developed a structured framework to classify and measure the commercial value derived from their agents, using a multi-model stack that combines capabilities from Mistral, Anthropic, and Cohere according to the specific task. LangChain acts as the orchestration layer that allows these diverse agents to be deployed in a unified manner. Their examples include agents that handle contract renewals, sentiment analysis, or internal ML tools—each requiring different types of reasoning and information access. JP Morgan Chase, meanwhile, implemented a supervisor architecture in LangGraph where human participation is not optional but fundamental to system reliability. They have designed specific sub-agents for tasks like document search using MongoDB, and use LLMs as "judges" that evaluate decisions within the workflow as a reflection step before executing actions.

The most interesting convergence between both cases lies in their critical dependence on specialized ML models and database access tools, but above all in that both systems prioritize information quality and precision over response speed. Cisco emphasizes the importance of experimentation teams and "fast failure" as a development methodology, while JP Morgan Chase adopts the philosophy of "start simple and refactor frequently," defining clear metrics and using continuous evaluation-based development. Both approaches recognize that production agents cannot be experimental systems—they require robustness, observability, and crucially, the ability to explain their decisions to humans who maintain final responsibility for outcomes.
‍
Agents as Developers
‍
The presentations by Michele Catasta from Replit and Russell Kaplan from Cognition explored how these systems are transforming software development itself. Here, agents don't just process information but directly participate in writing, reviewing, and optimizing code, collaborating with developers in the creative process of software construction.

Michele shared Replit's experience in developing agents for full-stack code generation, systems capable of working fluidly both on new projects and connecting to existing repositories. Their approach prioritizes creating highly collaborative environments where developers and agents can work together transparently, always maintaining security as a fundamental principle. The platform allows agents to maintain continuity with previous work, adapting to project context and developer preferences. Russell, meanwhile, demonstrated how Cognition has developed specialized agents that work specifically with source code, focusing on complex contexts and very specific programming languages. Cognition's approach also emphasizes close collaboration between humans and agents, but from a more specialized perspective, where agents can understand existing architectures and make precise modifications that respect system conventions and limitations.

What emerges from both experiences is that they have had to face the same fundamental technical challenge: managing massive context for agent memory. An agent working with code needs to maintain in memory not only the immediate task, but potentially thousands of lines of code, dependencies, documentation, change histories, and project-specific conventions. Both Replit and Cognition have developed significant innovations in data flow and processing, creating memory architectures that go far beyond traditional context windows. Both bet on highly collaborative environments that require agents to be able to explain their decisions, maintain transparency in their processes, and guarantee security in every interaction—recognizing that in software development, trust between human and agent is fundamental for adoption and system success.

Challenges and Perspectives
‍
The talks by Shreya Shankar from UC Berkeley and Andrew Ng provided valuable insights from two angles: research and experience. Shreya shared lessons on building reliable agents for data processing, based on her research for developing specialized tools for large-scale document analysis. Her work illustrates a fundamental challenge: when working with large volumes of complex information, obtaining precise results depends not only on formulating the right questions, but on the system being able to automatically optimize how to process and analyze all that information. What seems simple in theory—giving clear instructions—becomes very complex in practice when you need to maintain precision at massive scale.

Andrew complemented with viewpoints derived from his practical experience with agents. His key points revealed important opportunities: agents can be effectively integrated into both linear and non-linear workflows, evaluation of agentic systems must be considered from the design's inception, and emerging protocols like MCP and A2A promise to significantly expand capabilities (though he acknowledged they are still very recent). One of the most interesting observations was that the main skill for effectively using these tools will be fundamentally communicative—telling a computer exactly what is needed in a specific domain.

This perspective contrasts fascinatingly with Shreya's findings, who demonstrated that the greatest potential of an agent is obtained when the initial instruction is adequate, but the system can automatically optimize how to execute that instruction. Both perspectives converge in highlighting the importance of human involvement and expertise in the process and functioning of agent systems. Whether designing pipelines that automatically optimize themselves or formulating precise instructions that maximize agent effectiveness, the human factor emerges as the critical determinant of success, requiring collaboration between technical expertise and communication skills.

Final thoughts
‍
The technological maturity of AI agents is undeniable. These production implementations demonstrate that generative AI has evolved from experimental promise to functional tool, with systems that generate measurable value and transform fundamental processes. The versatility of the five-pillar architecture—orchestration, memory, tools, human-in-the-loop, and observability—provides a robust framework that transcends specific industries.

However, this technological maturity also reveals the enormous unexplored potential in sectors that could deeply benefit from these systems. Academic research could leverage agents to accelerate scientific discoveries and interdisciplinary collaboration. Environmental applications could use these systems for ecosystem monitoring and resource optimization. The public sector could implement agents to improve citizen services or control mechanisms. Social organizations could utilize this technology to plan community initiatives in pursuit of welfare and social justice.

The lessons about human-agent collaboration, the importance of clear instructions combined with automatic optimization, and the need for transparent and reliable systems are universally applicable. The technology is ready; what we need now is expansion of imagination and application. The future of AI agents will be defined not only by technical advances, but by our collective capacity to imagine and build applications that address the most diverse challenges of our society. The opportunity is immense, and the moment is now.‍‍