OpenAI Just Dropped Symphony: The First AI That Actually Works

OpenAI has released Symphony, a system that deploys AI agents to perform real coding jobs autonomously. Instead of merely assisting developers, Symphony allows AI agents to take on tasks directly from an issue tracker, such as fixing bugs or building features. The system monitors task statuses, and when a task is marked 'ready for agent,' Symphony automatically activates an AI agent to begin work.

Upon activation, Symphony initiates an 'implementation run,' where the AI agent attempts to complete the task from start to finish. A separate, isolated workspace is created for each task, preventing the AI from accidentally affecting other parts of the project. The AI reads the task description and proceeds to write code within this secure environment.

Symphony enforces a 'proof of work' requirement before any code is accepted. This includes running automated tests, generating CI reports, passing unit tests, and producing a walkthrough of the changes made. Only after all these checks are successful does Symphony proceed to 'landing,' where the AI submits a pull request, mirroring a human developer's workflow. Instructions for the AI are stored in a workflow.md file within the code repository, allowing version control of the AI's behavior alongside the code itself.

Xiaomi has introduced MClaw, an AI agent integrated directly into the phone's operating system, offering capabilities far beyond typical AI assistants. MClaw operates at the system level, granting it access to the phone's apps, settings, and connected devices within the extensive Xiaomi ecosystem, which includes over 1 billion devices. The AI is powered by Xiaomi's Mimo large model, developed by their AI team.

Users can give MClaw high-level instructions, such as "Prepare the house for my friend in 30 minutes," and the AI will autonomously execute a series of actions like adjusting lights, opening curtains, and changing air conditioner temperatures. MClaw functions using an inference execution cycle, where it receives an instruction, selects from over 50 system-level tools (e.g., launching apps, adjusting settings), executes a tool, analyzes the result, and then decides the next step. Users can observe this process in real-time.

To maintain task coherence, MClaw employs a three-level context memory system, ensuring the AI remembers its original goal even through multi-step tasks. It also features personal context understanding, reading information from messages, calendars, and usage patterns to proactively manage user needs. For example, it can update calendars from train ticket messages or recommend subscription cancellations based on spending analysis. Privacy is maintained by processing most data locally on the device, with sensitive actions requiring user confirmation.

Microsoft has released 54 Reasoning Vision 15B, a multimodal AI designed for efficiency and power, understanding both images and text. Unlike larger models, this compact 15-billion parameter model focuses on strong multimodal reasoning without the heavy computational demands. It combines the 54 reasoning language model with a SIGLIP 2 vision encoder, using a 'MIDI fusion' design where the vision encoder converts images into tokens for the language model to process alongside text.

Training involved approximately 200 billion multimodal tokens, building on earlier stages. Microsoft emphasizes that multimodal AI often fails due to perception issues rather than weak reasoning. To address this, the model uses a dynamic resolution vision encoder supporting up to 3,600 visual tokens, enabling it to analyze complex screenshots, documents, charts, and graphical user interfaces effectively.

The model incorporates 'mixed reasoning training,' where about 20% of training data includes 'think tags' to teach structured reasoning for complex problems. The remaining data focuses on perception tasks like image captioning and OCR. This approach allows the model to respond quickly when reasoning isn't needed but perform structured reasoning when required. It excels in scientific/mathematical reasoning over visual information and in computer use agents, interpreting screen content to automate actions.

OpenAI's Symphony is built using Elixir and the Erlang Beam runtime, chosen for its reliability in handling numerous processes and recovering from failures. This allows Symphony to manage hundreds of AI coding tasks concurrently without system-wide crashes if an individual agent fails. The system tracks everything using PostgreSQL through Ecto and runs continuously as a daemon.

OpenAI highlights the importance of 'harness engineering' for AI agents to function effectively within a codebase. This means the project repository must be structured in a machine-understandable way. Key requirements include locally and reliably runnable tests without external dependencies, machine-readable documentation, and a modular code architecture that allows agents to modify parts without breaking the entire system.

Symphony's scope is focused; it acts as a scheduler, runner, and tracker, sitting between project management tools and the codebase. Its primary role is to dispatch AI agents to tasks and manage their execution from start to finish. This specialized function underscores the need for well-prepared codebases to maximize AI agent utility.