Jean-Denis Greze reframes agent-to-agent collaboration as a search problem. The decisive model call needs the right context, but no trustworthy system can simply grant one agent permanent access to every person's email, company record and private workspace.
He compares five approaches: broad access inside a trust boundary, privacy-preserving custom tools, shared company knowledge spaces, human-mediated requests and a controlled black-box agent that identifies the minimum people whose approval is required. Each approach trades automation against privacy, security and operational complexity.
Greze sees the clearest near-term value in AI-maintained shared knowledge. A sweeper agent can propose or apply policy-governed updates from private silos to a company wiki, allowing useful context to accumulate without making every source universally visible.
The risks include prompt injection, poisoned shared memory, accidental disclosure and unclear audit responsibility. His recommended direction is an auto mode that handles low-sensitivity information, escalates uncertain cases to people and becomes more capable as policy enforcement improves. Product promotion and event logistics are omitted.
Watch on YouTube
