temp srija, neha, sheona

Inspiration

Building a product with AI should be accessible to everyone, not just experienced developers. Today, non-technical founders can use AI coding tools to build software, but turning an idea into a working product still requires managing tasks, debugging failures, coordinating tools, and figuring out whether the code actually works.

We wanted to make building with AI feel less like babysitting a collection of bots and more like working with a real team.

This project brings that idea to life through a virtual office where AI agents have distinct responsibilities, coordinate their work, and provide evidence of their progress. Our goal is to help anyone with an idea take the first step toward building it.

What it does

This project turns a project into an AI-powered development team.

Users describe what they want to build, and a manager agent breaks the goal into tasks and delegates work to specialist agents. Users can watch progress in a 3D office, inspect individual agents, and understand what is happening without constantly reading terminal logs.

Our core features include:

• Task delegation: A manager agent coordinates specialists responsible for research, planning, testing, and implementation.
• Agent communication: Agents pass relevant findings directly to one another, with messages tracked through delivery and acknowledgment.
• Coordinated file access: Tasks that affect the same files are sequenced to reduce conflicting changes.
• Evidence-based completion: Tasks require supporting evidence, such as observed test results, code diffs, or screenshots, before they can be marked complete.
• Human oversight: Users can inspect progress, review work, and intervene when an agent becomes stuck.

We want to make AI-powered development more understandable, accessible, and trustworthy for first-time builders.

How we built it

This project combines agent orchestration with an interactive 3D workspace.

We use Three.js to render the virtual office and its agents, with animations reflecting their current task states. Blender and MCP tools help create the office environment.

Codex powers the development agents through isolated sessions. A manager coordinates tasks, while custom tools enforce delegation, status updates, communication, acknowledgment, and review workflows.

A local server handles persistence, replay, and Git checkpoints. Our evidence gate connects task completion to observable outputs rather than relying solely on an agent's own claims.

Our technology stack includes TypeScript, JavaScript, Node.js, Three.js, Python, Blender, MCP, and Codex.

Challenges we ran into

Verifying that work was actually completed. An agent can claim that tests passed without having run them. We built an evidence-based workflow that requires observable outputs before tasks can close.

Coordinating agents safely. Multiple agents working on the same files can overwrite each other's changes. We introduced declared file scopes and a coordination process to sequence conflicting tasks.

Replacing stalled agents. Restarting work too early can cause two agents to compete for the same task. We designed a handoff process that preserves context before a replacement begins.

Making the experience approachable. Terminal logs can make multi-agent workflows difficult to understand. The virtual office gives users a more visual way to follow progress, inspect work, and identify blockers.

Accomplishments that we're proud of

• Built a functioning virtual office for coordinated AI agents.
• Developed a manager-and-specialist workflow with direct communication and task handoffs.
• Created an evidence gate that makes it harder for agents to mark unfinished work as complete.
• Designed a coordination process to reduce conflicting file edits and preserve context when agents are replaced.
• Incorporated accessibility features, including reduced motion, keyboard alternatives, and status labels that do not rely on color alone.

What we learned

We learned that getting AI agents to perform tasks is only part of the challenge. Coordinating their work, preserving context, handling failures, and independently verifying results are essential to building a reliable system.

We also learned that interface design matters. Making an agent's state, responsibilities, and blockers visible helps users understand what is happening and where their attention is needed.

Most importantly, we see an opportunity to make AI-powered development more approachable for people who have ideas but lack the technical experience to build them independently.

What's next for this project

We want to make this project a practical AI development partner for non-technical founders and first-time builders.

Our next steps include:

• Idea-to-product workflows: Turn plain-English ideas into structured plans, features, and implementation tasks.
• Founder-friendly controls: Let users clarify requirements, approve plans, and make decisions without needing to understand every line of code.
• Stronger testing and verification: Expand automated checks and make evidence easier to inspect.
• Persistent project context: Preserve requirements, decisions, and outstanding tasks across sessions.
• Local-first AI: Explore on-device models and local execution to give users greater control over their data and AI workflows.
• Shared workspaces: Let teams collaborate while observing and managing their agents together.

Our vision is to make building software with AI feel less like managing a collection of disconnected tools and more like leading a team that helps turn ideas into reality.

Built With

• python