AI Agents Take the Wheel: Devin, SIMA, Figure 01 and The Future of Jobs

Three new agent-style AI systems—Cognition AI’s Devin, Google DeepMind’s SIMA, and Figure 01—signal a shift from chatbots that describe work to systems that can carry it out in constrained environments. The common thread is that each system acts like a “container” around a vision-language model: swap in a stronger underlying model (GPT-5 or Gemini 2, for example) and performance can jump quickly, often overnight. That matters because it reframes the pace of progress—from slow, incremental model training to rapid capability upgrades delivered through the same agent shell.

Devin is positioned as an AI software engineer built around a code editor plus browser access. It doesn’t just interpret instructions; it plans steps, reads documentation, and executes changes autonomously. In a short demonstration, it reads a blog post, runs the referenced code, and then iterates—finding edge cases and bugs the original instructions didn’t mention—before returning final results and even bonus images. The headline metric comes from SWE-bench, a benchmark built from 2,294 real-world software engineering problems drawn from historical GitHub issues and their accepted solutions. Devin scores nearly 14% on the subset it was tested on, far above Claude 2 and GPT 4 at about 1.7%. But the benchmark design also shapes what “success” looks like: it emphasizes issues with solutions that introduce new tests, which may underrepresent harder bugs where writing a clear test is difficult. Even so, the transcript argues that Devin’s score could rise sharply with a stronger multimodal, long-context model—especially because SWE-bench tasks often require locating relevant code inside large repositories and coordinating edits across multiple files.

Google DeepMind’s SIMA pushes the same agent idea into games. SIMA uses mouse-and-keyboard control with pixel input, trained on data gathered from humans playing a wide range of games, including commercial titles and Google-made environments. The key finding is positive transfer: training across many games helps SIMA perform better on new games than agents specialized for a single environment. In some cases, the generalist approach even beats the specialist trained on that exact game, and performance is described as approaching human capability in the tested settings. The transcript links this to broader robotics lessons—co-training with skills from other platforms can enable novel task performance.

Figure 01 brings the concept into the physical world. It’s described as a humanoid robot that takes multiple images per second and performs tasks like dish-related manipulation without teleoperation, with intelligence attributed to GPT-4 Vision-style perception. The estimated cost—roughly $30,000 to $150,000 per robot—still limits widespread deployment, but the company’s roadmap aims at automating manual labor broadly, potentially reducing labor costs toward the price of renting robots and eventually letting humans “leave the loop.”

Across all three systems, the transcript repeatedly returns to a job-market tension: capabilities are improving fast, yet timelines remain uncertain and control over deployment is contested. Still, the overall direction is clear—agents are becoming operational, and their performance is increasingly tied to the next generation of underlying multimodal models and tool-augmented reasoning.