Things I want to do
This page is June's roadmap for Zenbu: the direction I want to take, systems to improve, and projects to build next. These are goals, not claims that the work is finished or commitments to a release date.
1. Bring Coding Agent into a chat-like workspace
Combine Coding Agent's execution tools and orchestration with a chat-like application experience. I want to move between conversation and hands-on agent work naturally, with local and remote workflows usable interchangeably.
The goal is a coherent workspace where conversations, tasks, files and results remain connected as work moves between my machine and remote execution.
2. Make browser actions genuinely useful for agents
Improve browser interaction for real agent use cases: navigating applications, understanding page state, completing multi-step actions, and checking whether those actions actually succeeded. Browser work should be reliable enough to participate in larger workflows, with useful recovery when a page or action fails.
3. Add native connectors that work through normal user access
Give agents more first-party tools for different data sources without requiring users to provision a separate developer API integration for every service. Where supported, agents should be able to use the same signed-in access and user-facing workflows a person would use to find and work with their own data.
Keep provider-specific access behind native connector/tool boundaries so the agent receives useful, validated inputs and results rather than having to understand every service's transport or setup details.
4. Make the coding-agent harness faster and more effective
Optimize the harness for getting real work done: lower idle resource use, responsive interaction, efficient execution and context handling, and better coordination across tasks. Measure performance and completed-work outcomes, not just the number of agents or tool calls.
5. Mix local and remote models for better results
Combine local and remote model execution in the same workflow. Use their different strengths deliberately, with appropriate routing and handoffs, rather than making every task use one model or one execution location.
The goal is better results with sensible latency and resource use, while making it understandable which model and execution environment did the work.
6. Build a software factory for games
Develop repeatable agent workflows that take a game from an idea through code, assets, animation, playtesting, optimization and a playable release. The output should be a coherent game that works, not merely a collection of generated files.
An important subgoal is better MCP integration and better observable results from our open-source creative tools, especially Blender and Aseprite-compatible pixel-art tooling. Agents need to create, inspect and revise assets, preserve editable sources, and verify how the results look and behave inside the game.
Repository map
For the projects and libraries that exist today, see the repository overview, directory guide, and technical knowledge index.