Start the fixed demo scenario, let the agent research, then add one requirement while it works. The board revision changes and only the newly invalidated checks should return to the research queue.
Agent Utility Index · shared research board
Humans control the rules. Agents control the research.
Change a requirement while the agent works. The board keeps stable IDs and revisioned evidence cells, so the agent can re-check only what your change invalidated instead of restarting the whole comparison.
The webpage becomes shared working memory between the human and the agent.
The point is not to build another research chatbot. The human can change the decision directly in the browser while the agent can read and update that exact same state through WebMCP.
Add a new requirement on the page while the agent is working. The board revision changes, the agent reads the update, and continues only on the evidence that is now missing.
Agent creates the board
The AI turns the user's research goal into a shared decision board with explicit must-have requirements.
Agent researches normally
The AI can use its own search and browsing abilities, then writes candidates and source-backed evidence into the live page through WebMCP.
Human changes the page
The person can add or remove a requirement directly on the board instead of re-explaining the entire task in chat.
Agent reads the new state
The agent calls the WebMCP board tools, sees the new revision, and knows exactly which evidence checks are now missing.
Both see the same decision
Candidates, evidence, missing checks, and activity stay visible in one shared browser surface while the conversation continues.