OpenAI Is Back in the Game - For Now

Why o3 and o4-mini are different and how they relate to Gemini-2.5-Flash

This Week in AI Agents

Welcome to State of AI Agents and another wild week in AI , with several breakthrough developments pointing toward more capable, adaptive, and self-directed AI systems. Leading AI labs and researchers introduced powerful new AI models and frameworks that enable agents to reason more deeply, plan more flexibly, and even improve themselves over time. OpenAI unveiled its most advanced reasoning AI models to date alongside a long-awaited open-source agent, signaling a push toward AI that can autonomously use tools and handle complex, multi-step tasks. Google DeepMind responded with a novel “hybrid reasoning” model that lets developers toggle an AI’s depth of thinking on demand, balancing quick responses against deliberative problem-solving.

Beyond the tech giants, academic teams proposed innovative agent architectures addressing some of the field’s toughest challenges: how to give AI agents better memory and state-tracking for long tasks, how to make them more robust against getting stuck, and how multiple agents can collaborate or specialize to outperform single monolithic systems. Researchers also demonstrated an agent that doubles the performance of prior systems at web-based information extraction by converting its actions into reusable programs. Another team showed how an AI “doctor” agent can self-evolve its workflow—adding new subtasks or decision branches when it encounters mistakes—eventually outperforming human-designed workflows in medical diagnostics.

A new perspective from DeepMind’s pioneers argues that AI will truly excel when agents can learn continually from their own “streams” of experience in the world, rather than being limited to static training data or short interactions. In summary, this week’s research suggests that autonomous agents are becoming more powerful and independent, with advances that improve their reasoning abilities, practical utility, and capacity to operate (and even improve) with minimal human intervention. The implications range from safer, more general AI systems to concrete economic benefits as agents tackle complex tasks in domains like the web, software engineering, and healthcare. The following report delves into these developments, explaining what each contribution is and why it matters, from high-level innovations to technical details for expert readers.

Advanced Reasoning Models: Fueling Autonomy

OpenAI's Reasoning 2.0: "o3" and "o4-mini" (official blog)

OpenAI's introduction of two specialized reasoning models is much more important to the development of autonomous agent capabilities than their previous two releases (GPT-4.1 and 4.5). These models - codenamed "o3" and "o4-mini" - represent their most advanced reasoning systems to date, with built-in capabilities for chain-of-thought reasoning, tool use, and visual understanding. What sets them apart is their integration of reasoning and action in a unified architecture. Previous systems typically required external orchestration to decide when to use tools like web search or code execution. The o3 model demonstrates remarkable agency by proactively identifying when to deploy tools without explicit instruction. In one demonstration, when shown a physics poster from 2015, the model autonomously initiated a web search for newer research and compared the findings to the poster's content - all without being prompted to do so. This is a critical advancement in agent autonomy: the ability to independently recognize information gaps and take appropriate actions to fill them. The models also process visual inputs directly, allowing them to interpret sketches, charts, or photos as part of their reasoning process - enabling multimodal chains of perception, reasoning, and action.

Google's Gemini 2.5 Flash: Controlling the Depth of Thought (official blog)

Google DeepMind's introduction of Gemini 2.5 Flash brings a different but complementary innovation: controllable "hybrid reasoning." This system allows developers to toggle the model's intensive reasoning capabilities on or off and even set a specific "thinking budget" for each query. With reasoning disabled, the model provides fast responses for time-sensitive applications. When enabled, it engages in deeper step-by-step thinking for complex problems, performing second only to Google's most powerful models on challenging benchmarks. This approach reflects a pragmatic insight: not every task requires extensive deliberation. By giving developers control over the reasoning/speed tradeoff, Gemini 2.5 Flash enables more efficient resource allocation. This mirrors human cognition, where we reserve deep thinking for difficult problems while handling routine tasks with minimal cognitive effort. For practical applications, this means greater flexibility in balancing user experience, cost, and performance. A customer service agent built on this technology could respond instantly to simple queries while thoughtfully working through complex customer issues - all within a single system.

Going Deeper: The Latest Research Trends

Now that we’ve got the obvious must-know topics out of the way, let’s move on to some more intricate insights that will put you ahead of the information curve. This week, we will dive into the following research topics:

• Adaptive Planning

• Hierarchical State Management

• Breakthroughs in Specialized Agents

• Self-improvment

Thank you for reading this issue of State of AI Agents! Did you know that you can reimburse your subscription to LLM Watch as part of your company’s learning budget? Here’s a tutorial on how to get your invoice from Substack.