The Power of Meta Access in the Software of the Future --- Demonstrated on LLM Chat UIs and Chat AI
A paper proposes giving assistants structured, permissioned access to application state and capabilities. Its prototype supports bounded retrieval and confirmed, reversible changes, while identifying control-plane and prompt-injection risks.
The paper describes a ladder of **five meta-access levels**, from feature tutoring through settings access and retrieval to confirmed changes and full orchestration. Its Meta Chat prototype builds bounded context from summaries, settings, and relevant messages, and supports confirmed, reversible changes. For builders embedding agents in an IDE or other workspace, expose state and actions through a structured, permissioned interface. Keep control of the harness separate when an agent can alter its own settings, and require confirmation for mutations. This is a proposal grounded in one chat-client prototype, not evidence of a broadly tested interface pattern. The authors identify grounding, privacy, prompt injection, and trust as open risks.
No separate implication has been recorded yet.
The paper describes a ladder of **five meta-access levels**, from feature tutoring through settings access and retrieval to confirmed changes and full orchestration. Its Meta Chat prototype builds bounded context from summaries, settings, and relevant messages, and supports confirmed, reversible changes. For builders embedding agents in an IDE or other workspace, expose state and actions through a structured, permissioned interface. Keep control of the harness separate when an agent can alter its own settings, and require confirmation for mutations. This is a proposal grounded in one chat-client prototype, not evidence of a broadly tested interface pattern. The authors identify grounding, privacy, prompt injection, and trust as open risks.
The paper proposes a graduated interface for agent access to workspace state and actions, with confirmation for reversible changes and separate control of the harness. It makes permissions and user approval part of the interface design, but its single prototype does not establish that the pattern resolves grounding, privacy, or prompt-injection risks.