IBM partners with Anthropic, plus OpenAI drops AgentKit

The Rise of Visual AI Agent Builders

OpenAI’s release of AgentKit signifies a strong push towards making agent creation more accessible, particularly for non-technical users. However, it’s not the first of its kind, joining a growing ecosystem of low-code and no-code platforms like Langflow (now part of IBM), Crew AI, and LangGraph.

The core debate around these visual builders centers on their scalability and place in the development landscape:

• The Case for Visual Builders: Proponents argue that these tools democratize AI by allowing anyone to create agents through intuitive, “clicky, clicky” interfaces. They are ideal for connecting to common services (e.g., Google Mail, Google Sheets) and implementing enterprise necessities like guardrails, which can be added as simple blocks. Chris Hay notes that OpenAI’s builder is cleverly built on their Agent SDK, allowing users to start with a visual workflow and then export the underlying Python or TypeScript code for version control, offering a bridge between simplicity and professional development needs. They even simplify complex configurations by using an LLM to generate necessary schemas from natural language descriptions.

• The Critique of Visual Builders: Critics, like Mihai Criveti, draw parallels to past attempts at visual programming like UML. While attractive for getting started and for simple, one-off automations, these platforms often become unwieldy when projects scale. Challenges in version control, integration with legacy systems, and collaboration among multiple developers become significant hurdles. For building robust, enterprise-grade applications, writing code remains the more powerful and manageable approach.

Ultimately, the strategy for a player like OpenAI is to leverage its massive volume of users. By providing a good user experience for simple tasks, they can foster “citizen automation” that eventually finds its way into enterprise settings, similar to the trajectory of tools like Microsoft’s Power Automate.

Enterprise AI: The IBM & Anthropic Partnership and the ADLC

The strategic partnership between IBM and Anthropic highlights a maturation in the AI industry, moving beyond a focus on model capabilities to establishing the business processes required for enterprise adoption. A key output of this collaboration is a guide for an Agent Development Lifecycle (ADLC), a structured methodology for securely deploying enterprise AI agents.

The ADLC adapts the principles of the traditional Software Development Lifecycle (SDLC) for the probabilistic nature of AI. Key considerations include:

• Structured Governance: The ADLC provides a formal process for planning, coding, building, testing, releasing, deploying, operating, and monitoring AI agents. This is a critical requirement for enterprises before allowing agents to interact with production systems.
• Emphasis on Evaluation (Evals): A core component is the continuous evaluation of agent outcomes to ensure correctness. This can happen in real-time as the agent executes (allowing for retries on errors) or as a post-mortem analysis of many executions to gauge overall accuracy.
• An Evolving Standard: The guide is considered a living document, built on collaborative experience from IBM’s consulting, technology, and research divisions, along with Anthropic and real-world client engagements. It must evolve to address new challenges, such as the potential bias of using one LLM to evaluate another.

This partnership signals a broader market trend away from the “one model to rule them all” paradigm. As Olivia Buzek points out, LLM engineering is more complex than initially thought, and different use cases require different models—from small, efficient models for generating NPC backstories in video games to large, robust models for enterprise workflows. This diversity necessitates partnerships, as no single company can master all domains.

Fundamental Research: Understanding Modular Manifolds

Thinking Machines, a research lab from former OpenAI CTO Mira Murati, is taking a “first principles” approach to AI research, re-examining the fundamental assumptions of model training. Their work on modular manifolds addresses a core challenge in deep learning: training instability.

Chris Hay explains the concept with an analogy:

During training, a model’s weights are adjusted across many deep layers. On a “flat plane,” a small adjustment can send the model’s learning process wildly off course, causing a “gradient explosion” that ruins the training run. This is like pushing an astronaut in zero gravity—they fly off into space indefinitely.

A manifold introduces a “curvature” to this process. This curvature acts like gravity, keeping the weight adjustments within a stable range and preventing them from spiraling out of control. The astronaut is still pushed, but gravity keeps them within the planet’s orbit.

The practical implications of successfully using manifolds are significant:

• Increased Stability and Predictability: It makes the notoriously difficult training process more reliable.
• Reduced Costs: By preventing catastrophic failures in training runs that can cost millions of dollars, it makes AI development more efficient.
• Better, Faster AI: More stable training ultimately leads to more reliable and capable models, accelerating the pace of innovation.

AI in the Real World: The Case of Radiology

A recent article in Works in Progress titled “The Algorithm Will See You Now” highlights a fascinating paradox: despite years of predictions that computer vision would replace radiologists, the demand for and salary of human radiologists have soared. This prompts a deeper discussion on the true barriers to AI adoption in critical professions.

The consensus is that replacing a human expert involves far more than just pattern recognition.

• The Interface and Context Problem: An AI model lacks human senses and the ability to interact with patients and other doctors to gather crucial context. A radiologist’s job isn’t just to spot an anomaly but to integrate that finding with a patient’s history and communicate it effectively. Mihai Criveti argues that building an agent to replicate this data-gathering and communication process is a “huge engineering challenge.”
• Trust and Accountability: As a society, we have not settled the question of placing life-or-death decisions in the hands of machines. Olivia Buzek emphasizes that machine learning is fundamentally probabilistic, while human decision-making is not. An error rate will always exist. Furthermore, radiology involves interpretation and nuance that current systems cannot capture.
• Systemic and Security Risks: The real-world infrastructure of healthcare is a major obstacle. Systems running on outdated software are vulnerable to hackers, and vendor lock-in presents a risk of essential services being turned off. Rolling out a reliable, secure AI system across the entire healthcare industry would be a monumental, multi-decade effort.

While Chris Hay posits that a moral imperative will emerge to use AI once it is provably superior (using the analogy of choosing Stockfish over a human grandmaster to save the world), the panel largely agrees that the foreseeable future is a hybrid model. AI will serve as a powerful assistant for tasks like triage, measurement, and providing a “second opinion,” but the final accountability, communication, and contextual understanding will remain firmly in human hands.

• Tags:
• AI Agents
• AI Governance
• Feature
• Enterprise AI
• AI in Healthcare
• Modular Manifolds
• Agent Development Lifecycle
• LLM Engineering