CONCEPT Cited by 1 source
Elicitation gate¶
Definition¶
An elicitation gate is the agent-architecture mechanism by which an agentic system solicits approval, preferences, or missing information from the user or another system mid-execution, and — critically — cannot proceed past the gate without it. In the MCP / Agent Lee context, the elicitation is the confirmation prompt a user sees before a write operation executes on their behalf.
"Elicitations allow agentic systems to actively solicit information, preferences, or approvals from users or other systems mid-execution. When Agent Lee needs to take non-read actions on a user's behalf we use elicitations by requiring an explicit approval action in the user interface." — Source: sources/2026-04-15-cloudflare-introducing-agent-lee.
Gate vs prompt¶
The load-bearing distinction Cloudflare draws (Agent Lee post):
"The permission model is the enforcement layer, and the confirmation prompt you see is not a UX courtesy. It's the gate."
A prompt is a courtesy the agent can bypass (and under pressure will). A gate is a structural invariant — the agent cannot skip it by construction because approval is checked by a component the planner has no way to address or modify.
Agent Lee mechanism¶
The gate is implemented by a Durable Object that sits on the path from the sandbox to the upstream MCP server. The DO:
- Inspects the generated code (method + body).
- Classifies it read vs write.
- Proxies reads directly.
- Blocks writes until the user explicitly approves via the dashboard UI.
Because credentials live inside the DO (not in the generated code), even a jailbroken sandbox cannot route around the gate — it has no API key to make the write with. See patterns/credentialed-proxy-sandbox for the surrounding pattern.
Why it's necessary¶
An agent that can take arbitrary action on a user's account needs a boundary tighter than "the model is pretty reliable." Evals, hallucination scorers, and tool-call success rates (all of which Agent Lee runs) reduce the base-rate probability of an unintended write but cannot drive it to zero. The elicitation gate is the component that makes the probability of an un-approved write structurally zero.
Design shift: "An agent that can take action on your account needs to be reliable and secure. [...] These guardrails allow Agent Lee to truly be a partner alongside you in managing your resource safely."
Seen in¶
- sources/2026-04-15-cloudflare-introducing-agent-lee — the canonical in-wiki instance. "Agent Lee cannot skip this step." Surfaces the gate as an explicit architectural component of the MCP permission system, not a UX toast.
Related¶
- patterns/credentialed-proxy-sandbox — the architectural pattern that places the gate at a credential boundary the sandbox cannot cross.
- systems/model-context-protocol — the protocol layer whose write operations the gate authorises.
- systems/cloudflare-durable-objects — the runtime substrate hosting the gate in Cloudflare's deployment.
- systems/agent-lee — the production agent whose write path is fully gated.