• Updates

  • May's update expands our tool ecosystem with powerful new integrations. From TradingView charts to academic research via Arxiv, agents now have access to specialized tools that enhance their capabilities across finance, research, and development domains.

    Added

    • TradingView integration for financial chart generation

    • Arxiv search for incorporating academic research into projects

    • Test Four Protocol support for advanced strategies

    Improved

    • Tool selection algorithm picks optimal tools for each task

    • Parallel tool execution reduces overall task completion time

    • Better error recovery when individual tools fail

    Fixed

    • Resolved tool timeout issues affecting long-running operations

    • Fixed memory leaks in image generation pipeline

    • Corrected tool authentication errors in high-concurrency scenarios

  • April introduces our Unified Code Stream architecture, revolutionizing how agents handle complex, multi-technology projects. Generate full-stack applications with React frontend, Python backend, and smart contracts—all coordinated through a single execution flow.

    Added

    • Unified stream service coordinating multiple code generators

    • Support for mixed-technology projects (Web2 + Web3)

    • Intelligent dependency resolution across different languages

    Improved

    • Project scaffolding now includes proper file structure

    • Better handling of imports and cross-file dependencies

    • Sandpack preview supports more frameworks including Vue and Angular

    Fixed

    • Resolved file path conflicts in multi-technology projects

    • Fixed preview refresh issues with complex React applications

    • Corrected TypeScript compilation errors in generated code

  • March brings deeper integration with APRO Protocol's verified data feeds. Agents can now access real-time, cryptographically signed price data for better decision-making in DeFi projects. This integration ensures your generated smart contracts always use accurate, manipulation-resistant price feeds.

    Added

    • Direct access to APRO's verified BTC/USD price feeds•

    • Cryptographic signature validation for all price data•

    • Automatic fallback to multiple APRO nodes for reliability

    Improved

    • Smart contract templates now include APRO oracle integration

    • Price data caching reduces API calls by 60%

    • Better error handling for network disruptions

    Fixed

    • Resolved timestamp validation issues with APRO signatures

    • Fixed decimal precision errors in price conversions

    • Corrected rate limiting when accessing high-frequency data


  • Alongside our existing OpenAI GPT stack, we’ve added Anthropic Claude and Google Gemini, creating a three-engine hybrid architecture. Agents can now switch models on the fly, balancing speed, reasoning depth, and cost.

    Added

    • Claude 3 Sonnet – superior long-form reasoning and safety filtering

    • Gemini 2.5 Flash / Pro – fastest generation, 128 k-token context window

    • Dynamic Model Router – auto-assigns GPT, Claude, or Gemini based on task type

    Improved

    • Unified prompt templates for consistent cross-model output

    • Upgraded cost-management: average call cost down 28 %

    • Parallel multi-model inference cuts complex project completion time by 22 %

    Fixed

    • Resolved JSON parsing errors from model-specific schema differences

    • Patched fallback-retry logic under rate limits to prevent double billing

    • Tuned streaming buffers to eliminate lost output fragments

  • The January Update solves the critical preview synchronization issue that prevented real-time code visualization. Our new Task Coordinator ensures seamless code accumulation across HTML, CSS, and JavaScript tasks, finally making the preview panel truly interactive and responsive to agent outputs.

    Added

    • Task Coordinator for managing multi-step code generation projects

    • Real-time preview updates that accumulate code across sequential tasks

    • Code extraction service supporting multiple programming languages

    Improved

    • Preview panel now maintains state across all task executions

    • CSS extraction regex patterns handle outputs over 7000 characters

    • Context persistence prevents loss between related tasks

    Fixed

    • Resolved preview showing only first task output instead of accumulated code

    • Fixed CSS code extraction failures despite successful generation

    • Corrected run_id overwriting bug that broke task context sharing