Research

Research Summary

This research initiative explores how autonomous coding agents transform software development when they are used on real repositories rather than toy prompts. The scope includes Claude Code, Codex, OpenCode, OpenHands, Pi, Cursor, Windsurf, and other agent harnesses that can inspect code, call tools, make edits, run tests, and participate in review loops.

The current framing is informed by OpenSymphony, Claude Code practice, Codex workflows, Cursor automations, and articles such as From Spec-Driven Work to Work Orchestration. The central question is how to turn AI coding from a chat interaction into a repeatable engineering system.

Key areas of investigation include:

• Coding Harness Capability: Compare how agent tools handle repository navigation, edits, tests, shell commands, browser work, and review.
• Quality of Results: Evaluate correctness, maintainability, security, and architectural fit of generated changes.
• Human Review Loops: Study how planning, checkpoints, comments, and pull requests improve agent output.
• Model and Tool Flexibility: Compare provider and harness switching across tasks that require different context, latency, and cost profiles.
• Engineering Reliability: Track recovery, reproducibility, state management, and handoff quality across longer work sessions.

By addressing these areas, we aim to identify the practices, tools, and evaluation methods that make AI-driven software development dependable enough for production engineering teams.

Research Summary

This research investigates personal agent runtimes: systems that operate on behalf of a user inside local, remote, desktop, browser, or chat-controlled environments. The scope includes OpenClaw, Hermes, GasTown, OpenAI Operator, Manus and OpenManus, browser-use patterns, and other runtimes that combine tools, memory, permissions, and execution state.

We explore generalist agents like OpenAI Operator and Manus, open-source counterparts such as OpenManus, browser automation frameworks such as Browser Use, and local/remote agent gateways that can coordinate with IDEs, terminals, calendars, chat systems, and web browsers.

The objective is to analyze the architecture, robustness, security model, user experience, and operational burden of these runtimes so teams can decide when personal agents should run locally, remotely, inside an IDE, behind a gateway, or as part of a larger automation platform.

Research Summary

This research focuses on the design, comparison, and evaluation of communication and coordination strategies for multi-agent systems (MAS), tackling the challenge of interoperability between heterogeneous agents and frameworks.

We analyze diverse intra-framework architectural patterns (e.g., graph-based orchestration, agent-centric messaging, structured teams) and investigate emerging inter-framework open standards critical for breaking down silos. Key protocols studied include A2A (Agent-to-Agent) for task-oriented communication and discovery, MCP (Model Context Protocol) for standardized tool/data access, and OGP (Open Gateway Protocol) for signed gateway-to-gateway federation, approval flows, and controlled relay into local agent systems.

The research aims to evaluate the effectiveness of these strategies and protocols in enabling reliable knowledge exchange, dynamic task allocation, efficient coordination, and robust error handling, considering aspects like protocol compliance, cost, and reliability as highlighted by MAS benchmark gaps. We also explore adoption hurdles and security considerations for these evolving coordination frameworks.

Research Summary

This research surveys the infrastructure and governance patterns needed to run AI agents safely. As agents interact with codebases, execute commands, access credentials, browse the web, and call enterprise tools, the boundary between productivity and risk becomes an architectural concern rather than a policy footnote.

We investigate containerization, virtualization, tenant isolation, remote execution, credential brokering, approval flows, and specialized cloud platforms like E2B. We also explore standardized development environment platforms such as Daytona that enable consistent and rapidly provisioned workspaces for agent execution.

The goal is to evaluate the capabilities, limitations, security guarantees, and integration patterns of these techniques across coding agents, personal agent runtimes, browser agents, and enterprise automations.

Research Summary

This research tracks the fast-moving model strategy landscape with a practical enterprise lens. Rather than treating model development as one planned distillation project, the current work compares open-weight and closed models, local and hosted inference, training and retrieval, and single-model versus routed systems.

Key areas of investigation include:

• Open Model Selection: Track Qwen, Gemma, Kimi, DeepSeek, OLMo, Granite, Nemotron, and other open or local candidates for coding, reasoning, multimodal, and workflow tasks.
• Training vs. Retrieval: Evaluate when fine-tuning, adapters, synthetic data, retrieval, context engineering, or hybrid inference produce better results.
• Model Routing Economics: Compare cost, latency, privacy, reliability, and quality when routing tasks across local, open, and frontier models.
• Architecture Watch: Continue tracking architectures such as diffusion language models, Mamba-style hybrids, Titans-style memory, long-context models, and multimodal-native small models when they affect deployability.

The goal is to help teams choose the least expensive model strategy that is still good enough for the task, while preserving a path to train or adapt models when the evidence supports it.

Research Summary

This research investigates the G-Eval framework, leveraging advanced LLMs as judges for robust, reference-free evaluation of language tasks. A core focus is the development of high-quality, diverse synthetic evaluation datasets, generated using techniques like diverse prompting and role-play to create varied inputs (prompts, questions) without relying on reference outputs.

We explore how G-Eval utilizes LLMs’ reasoning capabilities to establish evaluation criteria and perform nuanced, multi-dimensional assessments that show stronger correlation with human judgment than traditional metrics. The integration of these synthetic datasets into evaluation pipelines is examined, alongside critical analysis of challenges such as potential evaluator bias and reliability, and strategies for mitigation.

Research Summary

This research addresses critical gaps in current agent benchmarks, which often overlook operational reliability, step completion, recovery, cost, and human handoff. To enable more rigorous evaluation and optimize human-AI teamwork, we propose and investigate a Dual-Layer Evaluation Framework. This framework distinctly assesses:

• (A) Individual Specialist Agent Competence (accuracy, efficiency, robustness, protocol interaction) in isolated tasks,
• (B) System-Level Coordination (task routing, workflow efficiency, communication effectiveness, error recovery, overall task success) in complex, collaborative scenarios.

This approach aims to provide deeper insights into both component strengths and emergent system behaviors.

Research Summary

This research initiative focuses on AI media production workflows, including narrated article videos, text-to-video generation, image generation, music generation, and reusable tooling for turning source material into polished media assets.

Key activities will include:

• Comparative Model Analysis: We will systematically evaluate image, music, and text-to-video models, assessing their strengths, weaknesses, and suitability for various applications.
• Technical Specification Tracking: We will maintain a detailed record of technical specifications for key models. This includes parameters such as output resolution, frame rates, supported styles, and control mechanisms.
• Results Curation and Evaluation: A significant part of this research will involve curating galleries of video outputs. These galleries will be used to compare models based on critical factors like:
  • Prompt Adherence: How faithfully the video output reflects the input text prompt.
  • Reasoning and Coherence: The logical consistency of depicted actions and scenes, and the model’s ability to infer and represent complex relationships.
  • Text Capabilities: The quality and accuracy of any text rendered within the generated videos.

By systematically investigating these aspects, we aim to provide valuable insights into the current capabilities and limitations of AI-generated media and the tooling required to make it production-useful.

Research Summary

This research investigates real-time multimodal AI, focusing on seamless, context-aware interactions powered by live audio and video streaming. We analyze capabilities offered by platforms like the Gemini Multimodal Live API, which enable the development of sophisticated “ambient agents” capable of understanding and responding to their environment continuously (e.g., Ephor Coach, Highlight). A key focus is exploring applications of audio/video streaming and interaction specifically for Extended Reality (XR) environments, enhancing collaboration and contextual awareness across devices.

Research Summary

This research explores AI-powered methods for generating engaging and adaptive educational visuals on-demand. We investigate two primary approaches:

1. Utilizing Large Language Models (LLMs) to generate code for specialized visualization libraries (e.g., TikZ, Manim, p5.js) to create precise diagrams, animations, and interactive elements; and
2. Employing state-of-the-art multimodal generative models (like GPT-4o) for direct image creation from text prompts, suitable for illustrations and assets.

The project evaluates the effectiveness, feasibility, and integration strategies of these tool-based and direct generation techniques to enhance learning experiences and retention.

Research Summary

This research investigates how AI Tutors can transform education through personalized and self-directed learning. By leveraging NLP, ML, and adaptive algorithms, AI tutors customize learning pathways, content delivery (including dynamic text and image generation), and feedback in real-time. Our focus is on building and evaluating prototypes featuring these capabilities, including adaptive assessment modules, to enhance learner autonomy, engagement, and knowledge retention. We analyze the impact of these AI-driven approaches on learning outcomes and explore the technical and pedagogical considerations for effective implementation.