THIS WEEK IN AGENTS

The latest products, partnerships & predictions

• Resend CEO: How designing for AI agents is reshaping developer tools and email

• Pokémon No-Go: Claude’s advanced AI struggles to navigate Pokémon Red despite 3.7 upgrade

• How enterprises are balancing agentic automation with human oversight

• Swedish startup creates robot dog that learns like animals, not algorithms

• Got agency? Deloitte launches Zora AI platform with Nvidia to automate complex business tasks

• Keeping it real: 5 crucial business functions that should stay human in the AI era

• Manus challenges OpenAI’s Operator with autonomous AI agent for complex tasks

• Halliday raises $20M to build secure AI agents for blockchain networks

• Qualtric Control: Company launches AI Experience Agents to solve customer problems autonomously

• Mid-collar concerns: AI companies pivot to autonomous systems designed to replace human workers

What’s happening in agents right now

AI agents need better systems

The rise of autonomous AI agents represents perhaps the most significant shift in artificial intelligence since the emergence of large language models themselves. These agents, AI systems designed to independently execute complex tasks , are being deployed across numerous domains, from customer service to financial operations. But a clear theme is emerging: while we've made remarkable progress, we're still discovering fundamental limitations and challenges that need solving.

Looking at recent developments across the industry, it's becoming evident that effective agentic AI requires not just powerful models, but carefully designed systems, infrastructure, and operational frameworks to truly deliver on its promise.

The gap between promise and performance

Despite rapid advancements in AI capabilities, we're witnessing a reality check when it comes to autonomous agents. Take the recent experiment with Anthropic's Claude 3.7 Sonnet attempting to play Pokémon Red. Despite being one of the most sophisticated language models available, Claude struggles with basic navigation in the game world, frequently walking into walls and failing to interpret pixelated graphics effectively.

This experiment vividly illustrates the gap between theoretical AI capabilities and practical, autonomous functioning in even moderately complex environments. Claude's struggles highlight that we're still far from achieving truly autonomous agents capable of navigating unpredictable situations without human guidance.

Systems thinking for agent infrastructure

What's becoming increasingly clear is that successful agentic AI requires robust infrastructure specifically designed for agent operations. Zeno Rocha, CEO of Resend, points to this need by introducing the concept of "agent experience" as a progression from developer experience.

Rocha argues that APIs designed for AI agents must differ fundamentally from human-centered interfaces. This represents a profound shift in how we conceptualize software design, creating systems not just for human users but for autonomous AI collaborators.

Companies like Deloitte are already moving in this direction, having launched Zora AI, a platform featuring autonomous agents powered by Nvidia's technology. The platform has reportedly achieved impressive results in early deployments, including a 25% cost reduction and 40% productivity improvement in expense management.

This shift toward infrastructure for agents extends to blockchain networks as well. Halliday has secured $20 million in Series A funding to develop AI agents that can safely operate on blockchain networks. Their Agentic Workflow Protocol creates immutable safety guardrails, addressing critical challenges in AI-blockchain integration, a practical example of designing systems specifically for agent operations.

Learning from living systems

One of the most intriguing approaches to addressing the limitations of current AI agents comes from Swedish startup IntuiCell. The company has developed Luna, a robot dog with a digital nervous system that learns and adapts naturally like living organisms.

What makes this approach distinctive is that Luna learns through real-world interactions rather than pre-programmed responses or massive training datasets. The robot operates without pre-training, offline simulations, or extensive data centers, a fundamentally different approach from the data-hungry methods dominating AI development.

This biomimetic approach could offer solutions to one of the key challenges facing agentic AI: adaptability in unpredictable environments. By learning from how biological systems develop intelligence, IntuiCell's work points to alternative paths for creating truly autonomous agents.

Finding the human-AI balance

While the technical challenges are substantial, perhaps the most important consideration is determining the right balance between automation and human involvement. As agentic AI becomes more capable, businesses face strategic decisions about which functions should remain distinctly human.

McKinsey estimates that generative AI could optimize 70% of business processes by 2030, but successful implementation requires both technological advancement and organizational alignment. The key insight is that AI should augment human capabilities rather than replace human judgment entirely.

Business leaders should consider keeping five critical areas largely human: core message and values, thought leadership, strategic business decisions, client relationships, and creative direction. Overreliance on automation risks creating generic, forgettable companies that lose their competitive advantages.

Qualtrics' approach with their new AI-powered "Experience Agents" illustrates this balance in practice. These agents can autonomously resolve customer issues across multiple touchpoints, combining speed with empathetic engagement. The system aims to provide personalized responses to individual issues, potentially transforming customer service while maintaining human oversight through a dedicated AI ethics team.

Building better systems for an agentic future

As we move toward increasingly autonomous AI systems, the focus must shift from simply creating more powerful models to designing comprehensive systems that enable these agents to function effectively and safely in complex environments.

What remains unclear is how quickly these systems will mature. Will we see truly autonomous agents capable of handling complex tasks without human oversight in the next few years? Or will progress be more incremental, with human-AI collaboration remaining the optimal approach for the foreseeable future?

We publish daily research, playbooks, and deep industry data breakdowns. Learn More Here

Agents on the podcast

Our latest conversations around AI agents