Symbiosis

• 

  Human direction, agent execution

Inspiration

AI agents are becoming dramatically more capable. At the same time, the services we use every day are becoming increasingly agent-friendly through technologies like MCP.

That creates a new problem: as agents become capable of doing more work on our behalf, we need better ways to give them work, set priorities and boundaries, step in when human judgment is needed, and review what they have done.

Chat is great for giving an agent one task. It is less suited to overseeing an ongoing workload.

Then came WebMCP.

WebMCP gives us the tooling we need to make this relationship work particularly well.

What it does

Symbiosis is built around the symbiotic relationship between a human and their AI agent companion.

The human plans the work, prioritizes tasks, defines constraints, reviews results, and makes the decisions that require human judgment.

The agent works through that queue autonomously using the tools and capabilities it already has. It reports progress, records results, and pauses for a decision when necessary before continuing.

In short:

The agent executes. The human oversees, prioritizes, and decides.

Symbiosis makes that relationship visible in one shared workspace, with tasks moving between Backlog, Working, Needs Decision, and Done as the human and agent collaborate.

How we built it

Symbiosis itself was built collaboratively with AI, which felt particularly appropriate for the project.

I used ChatGPT to explore and challenge the original product idea, gstack to turn that concept into a detailed product and engineering design, and Codex to help implement and test the application.

A major design goal was keeping the architecture deliberately small. Symbiosis is the coordination layer - it does not try to duplicate the reasoning, web access, coding tools, email integrations, or other capabilities the agent already has.

Challenges we ran into

Our first iteration was built around a single workboard. It worked, but it quickly became clear that the model was too limited. Jobs could not be reused easily, workflows were difficult to compose, and the overall experience lacked the flexibility we wanted.

Through further iteration, we developed a three-level system of flows → workboards → jobs.

Flows orchestrate workboards, and workboards execute jobs.

That hierarchy gave us a much more elegant and reusable way to model agent workflows. Instead of rebuilding a process from scratch every time, users can define small reusable jobs, group them into workboards, and then combine those workboards into larger flows.

The challenge was finding a structure that stayed simple enough to understand while still being powerful enough to support loops, delays, schedules, human decision points, and more complex workflows.

Accomplishments that we're proud of

We are particularly proud of the flows → workboards → jobs system.

Its reusability is one of the most powerful parts of Symbiosis. Jobs can be defined once and reused across different workboards, while workboards can become reusable building blocks inside larger flows.

This turns agent interaction from a sequence of one-off prompts into something that can be structured, repeated, refined, and reused over time.

We are also proud that WebMCP feels fundamental to the product rather than being added simply for the hackathon. Symbiosis relies on WebMCP as the coordination layer between the human-owned workflow and the user's existing AI agent.

What we learned

The biggest thing we learned was where WebMCP is genuinely useful.

WebMCP works extremely well as a bridge between a conversational AI agent and a website. It allows the agent to understand structured application state, take actions, update progress, request decisions, and continue working while the human interacts naturally through chat.

We also learned what it is less suited to.

Trying to make the entire agent experience happen inside the webpage itself weakens one of WebMCP's biggest advantages: the agent already has its own context, identity, tools, and conversational interface.

The strongest pattern we found was to let the website act as the structured control surface while allowing the agent to remain where it already lives.

That separation became one of the core design principles behind Symbiosis.

What's next for Coop Queue

The immediate focus would be UI and UX refinement, clearer and richer tool outputs, and more advanced flow-control capabilities.

One particularly exciting area is intelligent conditional loops, where a flow can evaluate the result of a completed job and decide what should happen next based on that data.

We would also like to add better reporting, notifications, execution history, and visibility into long-running workflows.

Beyond that, the most important next step is user testing. We want to see how people actually use persistent agent workflows in practice, where they want more control, and where they are comfortable giving their agents more autonomy.

Built With

• chatgpt
• codex
• gstack
• react
• typescript
• vite