About AgentiCRE

Andrew Ng, a prominent figure in the field of Artificial Intelligence, recently made a public presentation where he likened AI to “the new electricity,” a general-purpose technology with the power to unlock countless new applications. While much of the recent buzz in AI has centered on foundational models and semiconductors, Ng emphasizes that the true value and most significant opportunities lie in the application layer. Among the myriad advancements in AI, Ng is most excited about a specific technical trend: agentic AI workflows, which he considers the single most important AI technology to watch.

Beyond Single-Shot Prompts: The Agentic Revolution

The typical way most users interact with large language models (LLMs) today is through “zero-shot prompting”, asking the AI to generate a complete output in one go, much like writing an essay from start to finish without backspacing. While LLMs perform surprisingly well in this manner, Ng highlights that humans don’t do their best work in this way, and neither do AIs.

This is where agentic workflows come in. According to Ng, an agentic workflow is an iterative process where an AI system breaks down a complex task into smaller steps, performs research, drafts, critiques its output, and revises, often looping through these stages multiple times. This approach, akin to a human thinking, researching, and revising, takes longer but results in significantly better and more robust outputs. For example, a legal team could use an agentic workflow to process complex documents, or a healthcare system could leverage it for diagnostic assistance.

The performance gains are stark. Ng cites the HumanEval Benchmark, which measures an LLM’s ability to solve coding puzzles. While GPT-3.5 achieved 48% accuracy and GPT-4 a much improved 67%, GPT-3.5 with an agentic workflow could achieve up to 95% accuracy. This demonstrates that the improvement from using an agentic workflow can “dwarf” the improvements from moving to a larger, more advanced foundational model alone.

The Four Pillars of Agentic Design

Ng identifies four major design patterns that builders are using to construct agentic workflows in their applications:

1. Reflection: This pattern involves an LLM critiquing its output. For instance, a “coder agent” might generate code and then be prompted to review and critique it, even suggesting improvements. This iterative self-correction, often combined with unit tests, significantly boosts performance from a baseline level.
2. Tool Use: LLMs are prompted to decide when to make API calls, such as searching the web, executing code, or performing specific actions like issuing refunds or sending emails. This expands the capabilities of agentic workflows by allowing LLMs to interact with external systems and data. Ng notes that LLMs are now often explicitly tuned to support tool use, which creates a higher ceiling for what agentic workflows can achieve.
3. Planning: For complex requests, an LLM can be prompted to devise a sequence of actions or steps to achieve the desired outcome. This allows the AI to break down a large task into manageable sub-tasks and execute them in a logical order, often involving different models or tools for each step.
4. Multi-Agent Collaboration: Instead of a single LLM performing all tasks, this pattern involves prompting an LLM to play different roles at different points in time, simulating multiple agents interacting with each other to solve a task. Ng draws an analogy to multi-threading on a CPU. While it’s still one processor, abstracting tasks into different “agents” can help developers break down and solve complex problems more effectively, often leading to significantly improved performance for various tasks.

The Dawn of Visual AI Agents

Ng is particularly excited about the rise of large multimodal models (LMMs), extending the power of agentic workflows beyond text to include image and video data. Just as with LLMs, LMMs can perform better with an iterative, step-by-step agentic approach compared to zero-shot prompting.

He demonstrated this with a “vision agent” capable of complex visual AI tasks. For example, by uploading an image of a soccer game and asking it to count players, the agent generated Python code to perform object detection and counting, then provided the accurate result. This capability allows businesses with vast amounts of visual data (images, videos) to extract significant value that was previously difficult to obtain. Other demonstrations included splitting video clips to identify specific events (such as a goal being scored) and generating metadata for video content, enabling searchable video databases. This significantly lowers the barrier to building complex visual AI applications.

Impact on the AI Stack and Development Practices

The emergence of agentic AI is not just changing how applications are built but also evolving the AI stack itself. Ng points to a new, emerging agentic orchestration layer (like LangChain or LangGraph) that makes it easier for developers to build these complex applications.

This rapid pace of development, driven by generative AI, means that while prototyping machine learning models has become much faster (from days to months), other parts of the software development process, such as product design, software integration, DevOps, and MLOps, still require time. However, the speed of ML prototyping is putting pressure on organizations to accelerate these other pieces as well. Ng advocates for a new mantra: “move fast and be responsible,” emphasizing the ability of smart teams to prototype quickly and evaluate robustly without shipping harmful products.

A critical bottleneck in this accelerated process is evaluations (evals). In the past, collecting test data was a small additional cost to training data. Now, with LLM-based apps often not requiring training data, collecting thousands of test examples becomes a significant bottleneck. Ng notes that building and collecting data are usually done in parallel rather than sequentially, and there’s still much innovation needed in how evals are built. He suggests that many teams delay systematic evaluations, but even quickly thrown-together, imperfect evaluations can be immensely helpful in complementing human judgment and incrementally improving systems.

Key Trends and Future Challenges

Ng identifies several crucial trends supporting the agentic AI revolution:

• Faster Token Generation: Agentic workflows require reading and generating a lot of text (tokens). Efforts to accelerate token generation through advancements in semiconductor and software technology will make agents significantly more efficient.
• LLMs Tuned for Tool Use: Modern LLMs are increasingly optimized not just for answering human queries but explicitly for supporting tool use and fitting into iterative agentic workflows.
• Rising Importance of Unstructured Data Engineering: With generative AI’s prowess in processing text, images, and video, managing and deploying unstructured data to create value is becoming a significant effort for businesses.
• Image Processing Revolution: While text processing is here, the image processing revolution is rapidly advancing, significantly increasing the range of possible applications.

He also touches on challenges and underrated areas:

• Bridging Business Needs to Agentic Workflows: It’s still difficult for businesses to break down existing processes into the right granularity of micro-tasks for agentic workflows. This requires a rare skill set to define steps, branches, and effective evaluations.
• Tactile Knowledge for Debugging: Building agents often requires “tactile knowledge” – the ability to quickly diagnose issues by looking at output traces and making informed decisions on what to do next. This skill is developed through practice with various AI tools and an understanding of their limitations.
• Underrated Opportunities:
  • Voice Stack Applications: Ng sees massive, often underestimated, opportunities for voice-based agentic applications in enterprises. While real-time voice can be challenging due to latency, agentic voice workflows offer more control and significantly reduce user friction, as people feel more comfortable speaking than typing for many applications.
  • AI-Assisted Coding: Ng firmly believes that AI-assisted coding makes developers significantly faster and that everyone should learn to code, as it enables better instruction to computers across all job functions. He dismisses the idea that AI will automate away coding jobs, comparing it to past fears that programming languages would become easier.
  • MCP (Model-Client Protocol): Ng sees MCP as a “fantastic first step” towards standardizing the interface for agents to plug into various tools and data sources, greatly simplifying the “plumbing” of data integrations. While early and still evolving, it promises to reduce the integration effort from N*M to N+M.
  • Agent-to-Agent Communication: Though very early, the concept of agents from different teams successfully interacting is a future frontier. Currently, multi-agent systems primarily work best within a single team due to the understanding of protocols.

In this presentation, Andrew Ng is extremely optimistic about the future of AI agents, believing they are expanding the realm of what is possible and opening up countless new applications. The ability to experiment and build faster than ever before makes this an exhilarating time for builders in the AI space.

 

In the first part of this series, we explored the critical “why now” behind the rise of the Frontier Firm and the phased journey organizations embark on to embrace AI. Now, let’s delve deeper into what truly defines these pioneering organizations. What makes a Frontier Firm stand out? How do they operate, and what unique traits set them apart in the rapidly evolving AI landscape?

The Core Identity: Intelligence on Tap, Human-Agent Teams, Agent Bosses

At its heart, a Frontier Firm is distinguished by its operational model, which is built on three foundational pillars:

1. Intelligence on Tap: This means that intelligent capabilities, powered by AI, are readily available, scalable, and on-demand throughout the organization. It’s no longer a scarce resource but an abundant utility.
2. Human-Agent Teams: The traditional human-only team is evolving into a hybrid model where AI agents work seamlessly alongside human colleagues. These aren’t just tools; they are integrated team members with specific roles and responsibilities.
3. Every Employee is an “Agent Boss”: This is perhaps the most significant shift. In a Frontier Firm, every individual, regardless of their role, learns to manage and leverage AI agents to amplify their impact. This involves building, delegating to, and overseeing AI-powered workflows.

This organizational blueprint enables Frontier Firms to scale rapidly, operate with unparalleled agility, and generate value at an accelerated pace. They are characterized by systems that are AI-operated but fundamentally human-led, striking a powerful balance between automation and human oversight.

Traits of Early Adopters: Spotting a Frontier Firm

While the concept of a Frontier Firm is still evolving, early adopters provide a clear glimpse into the future. These are the organizations already leading the charge, and they exhibit five distinct traits:

1. Organization-wide AI Deployment: AI isn’t confined to isolated departments; it’s integrated across the entire enterprise.
2. Advanced AI Maturity: They demonstrate a high level of sophistication in their pace of AI adoption, their mindset towards AI, their investment in AI, their broad adoption across the organization, and a clear understanding of AI’s return on investment (ROI).
3. Current Active Use of Agents: They are not just contemplating agents; they are actively deploying and utilizing them in live business processes.
4. Projected Moderate or Extensive Agent Integration Plans: They have concrete plans to expand agent integration significantly soon.
5. A Strong Belief that Agents are Key to Realizing ROI on AI Investments: They understand that agents are the primary vehicle for translating AI capabilities into tangible business value.

Evolving Workforce Dynamics: From Org Charts to Work Charts

The rise of the Frontier Firm isn’t just about technology; it’s profoundly reshaping the nature of work and the dynamics of the workforce.

Every Employee as an “Agent Boss”: The New Skill Set

The role of an “agent boss” is rapidly becoming a fundamental aspect of every job. This isn’t just for managers; it’s about individual contributors learning to build, delegate tasks to, and manage AI agents to amplify their personal and team impact. Interestingly, leaders are currently ahead of their employees in adopting this mindset, likely because working with agents plays to their existing strengths in delegating, guiding, and providing oversight.

The Shift from Org Chart to “Work Chart”

The traditional, static organizational chart, defined by functional silos, is becoming obsolete. Frontier Firms are moving towards a dynamic “Work Chart” model. This new structure is outcome-driven, with teams forming fluidly around specific goals and projects rather than rigid departmental lines. Agents expand the scope of individual employees, enabling lean, high-impact teams to assemble and disband as needed, much like a movie production crew. This enables companies to access the right talent and expertise on demand, without the need for constant, disruptive reorganizations.

Human-Agent Ratio: A Critical New Metric

Just as companies monitor human-to-human ratios in teams, the human-agent ratio is emerging as a crucial new business metric. Optimizing this balance is key to success.

• Balancing Act: Too few agents mean underutilized digital resources, while too many could overwhelm human judgment, leading to errors or burnout.
• Optimal Balance: The sweet spot is where agents significantly enhance productivity and foster innovation, while humans provide robust guidance, oversight, and strategic direction. This ensures efficiency without compromising quality or human insight.

The Emergence of “Intelligence Resources”

As AI agents become pervasive, a new organizational function is expected to emerge: “Intelligence Resources.” This new department will blend the capabilities of IT and HR, focusing on the strategic management of digital labor at an organizational level. It will be responsible for agent deployment, performance management, and ensuring the seamless integration of AI into human workflows, as well as maintaining ethical guidelines.

Strategic AI Integration: The Art of the Agent Boss

Effective AI integration in a Frontier Firm goes beyond simply deploying tools. It requires a strategic mindset and the development of specific skills across the workforce.

AI as a “Thought Partner”

To truly harness the power of agents, employees must adopt a “thought partner” mindset when interacting with AI. This means engaging in conversational exchanges, challenging AI’s outputs, brainstorming ideas, and sparking creativity, rather than simply issuing commands. It’s about a dynamic, iterative dialogue that leverages AI’s computational power with human intuition.

Key Skills for Agent Bosses

Becoming an effective agent boss requires a new set of critical skills:

1. Iterating with AI and Refining Instructions: The ability to provide clear, iterative feedback to AI models to improve outputs.
2. Knowing When and How to Delegate to AI: Understanding which tasks are best suited for AI and how to hand them off effectively.
3. Prompting with Context and Intent: Crafting clear, precise prompts that provide necessary context and define the desired outcome and data sources.
4. Refining Outputs Instead of Accepting First Drafts: Critical evaluation and a willingness to iterate on AI-generated content.
5. Critically Evaluating AI Outputs: The ability to spot weak reasoning, biases, or gaps in AI-generated information.
6. Knowing When to Push Back or Steer the Conversation/Plan: Understanding when to guide the AI, challenge its suggestions, or redirect its focus.

Strategic Agent Building and Management

Building and managing agents strategically is paramount:

• Build Strategically: Don’t just build agents for the sake of it. Identify specific business processes that need improvement, focusing on the “why” and the tangible benefits. Start with areas overwhelmed by data and build prompt-and-response agents to streamline information flow.
• Delegate Effectively: Calibrate the human-agent ratio based on the task’s complexity and sensitivity. Encourage agents to challenge assumptions by asking open-ended questions.
• Manage Performance: Just like human employees, agents need performance management. Upskill underperforming agents by refining instructions and providing accurate, up-to-date training materials. Set clear expectations and provide consistent feedback. Conduct “performance reviews” focused on measurable outcomes like productivity gains, cost savings, new revenue streams, or uncovered opportunities.

Human-Centric Approach: The True ROI

Ultimately, the real return on AI investment in a Frontier Firm comes from a human-centered approach. AI is a powerful enhancer, but humans remain uniquely capable of creativity, nuanced judgment, emotional intelligence, and building genuine relationships. The goal is to empower individuals to leverage these tools to elevate their work, not to diminish their role.

Data as the Foundation: Security and Responsibility

For AI to be truly effective, it must be grounded in the company’s proprietary data and knowledge. This ensures that AI is steered toward specific organizational goals and operates within the company’s unique context. This also underscores the critical importance of robust security, comprehensive data governance, and responsible AI practices. Without these foundational elements, the promises of the Frontier Firm cannot be fully realized.

The characteristics of a Frontier Firm paint a clear picture of the future of work. It’s a future where human ingenuity is amplified by intelligent machines, where organizational structures are dynamic, and where every individual is empowered to lead digital labor. Understanding these traits is the first step towards building an organization that not only survives but thrives in the AI era.

The Next Evolution of Software – and Why Real Estate Won’t Escape It

Commercial real estate, an industry historically slow to adapt to technological shifts, is now staring down a transformation it can’t afford to ignore. The catalyst is what Andrej Karpathy, a pioneer in modern AI, calls Software 3.0 , the new age of software defined by Large Language Models (LLMs) and autonomous, intelligent agents.

Karpathy’s framing is clear: LLMs aren’t just fancy chatbots or code-completion toys. They’re a new kind of programmable substrate, capable of ingesting mountains of unstructured data and performing reasoning-like tasks previously reserved for humans. This is not the next productivity app. It’s a wholesale shift in how software learns, acts, and scales.

For an industry like commercial real estate (CRE) fragmented, information-heavy, reliant on inefficient human workflows – the implications are as profound as they are unavoidable.

From Tools to Agents: What “Agent-Native” Really Means

In traditional computing (call it Software 1.0 and 2.0), we told the computer exactly what to do, line by painstaking line. In Software 3.0, the machine learns what to do by consuming data, millions of documents, conversations, contracts, leases, market reports, and responding dynamically to prompts and evolving contexts.

This shift makes software agent-native by design. Systems aren’t just static apps or dashboards; they’re digital coworkers, able to execute tasks with varying degrees of autonomy. Karpathy calls this the “autonomy slider.” Slide it down, and the agent works as a research assistant, summarizing trends, gathering data, surfacing insights on command. Slide it up, and the agent can orchestrate entire workflows: drafting LOIs, coordinating diligence, flagging risks, even negotiating terms, always under human supervision, but increasingly with fewer touchpoints.

The leap is not theoretical. It’s happening. And real estate is fertile ground.

Why CRE Needs Agentic Systems – Now

Why does this matter for commercial real estate? Because no other major asset class relies so heavily on imperfect information, human bottlenecks, and backward-looking assumptions.

Legacy underwriting still depends on static spreadsheets and the illusion of certainty in single-point estimates. Meanwhile, the reality is macroeconomic uncertainty, volatile capital flows, and local market asymmetries that no single analyst or broker can parse at scale.

Agent-native systems offer a structural advantage:

• Automation: Agents handle repetitive, rules-based tasks – scraping comps, checking capital availability, updating lease abstracts.
• Decision Support: By integrating live market data, demographic shifts, lending trends, and historical context, agents can forecast scenarios humans miss.
• Personalization: Agents can serve investors, tenants, and partners in tailored ways, from matching capital to opportunities to generating bespoke risk reports.

This is not the death of human expertise. It’s its amplification. Professionals stay in the loop but offload the drudgery.

The Infrastructure Catch: Why This Won’t Be Plug-and-Play

Karpathy is explicit: Software 3.0 demands new infrastructure. Faster GPUs alone won’t cut it. For real estate, which means:

• Robust, real-time data pipelines: agents are only as sharp as the data they’re fed. Outdated property records and stale comp sets won’t cut it.
• Agent orchestration platforms: when you have multiple agents (one underwriting, one managing lender communications, one tracking compliance) you need a secure layer to coordinate them.
• Ironclad compliance and security: real estate deals touch sensitive data, financials, and regulatory frameworks. AI agents must log their actions and prove their math.
• Human-agent collaboration interfaces: the goal is not to vanish humans but to elevate them. Interfaces must make it obvious what the agent knows, what it doesn’t, and how you can steer it.

If this sounds like a big lift, it is. But so was migrating brokerage listings online. So was accepting DocuSign. So was trusting third-party property managers with entire portfolios. And every time, the laggards paid the price.

No Straight Lines, But No Way Back

The path won’t be linear. LLMs still hallucinate. Data silos will resist integration. Some operators will cling to Excel until the last possible moment. But the trend line is unstoppable. The logic of agent-native systems is clear: the first movers who figure out how to harness autonomy safely, at the right points along Karpathy’s “slider”, will compound efficiency, insight, and competitive advantage.

The Real Work Starts Now

Software 3.0 isn’t a neat plug-and-play upgrade. It’s a mindset shift, a stack rebuild, and an invitation to think bigger about how we run, buy, build, and trade commercial real estate.

The question is no longer if the agentic era will arrive, but who will build the workflows, data channels, and governance structures to make it safe, compliant, and transformative.

Stay connected as we chart this frontier. Subscribe at agenticre.ai and join the conversation on how agent-native systems will reshape CRE for good.