WG: review agents-wg PR 22 (agent discovery SEP) with experimental/ext/agents as a second implementation #1238

Carried out of #1142. The implementation shipped in August. The outreach it gated has partly happened, and the target has moved.

What changed since this was filed

• The research PR (agents-wg PR 20) became a draft Extensions Track SEP, agents-wg PR 22 "MCP Agent Discovery Extension" (madhaviai, sponsored by LucaButBoring, opened 2026-08-20). Review happens there.
• The SEP answered the envelope question. It advertises under capabilities.extensions, as mcpkit does.
• The host bridge (agent: bridge a decoded agents/get result into an AgentSource #1144) and the Runner left mcpkit with the agent SDK extraction (caf24e8f). mcpkit holds the wire primitive and the Go client decoder only.
• mcpkit moved its extension ID to io.mcpkit/agents (7d66469f) until the shape is ratified.
• 2026-09-17: first comment posted. Review comment 4032278568 asked who enforces the "agents/get tools also appear in tools/list" rule. Head bfc43abb answered it the same day (servers SHOULD validate at configuration time, per-request responses MUST reflect the caller's authorization) and also listed "agent exposed as a single tool" as a rejected alternative.
• 2026-09-18: no runnable agents example remains in mcpkit. f24f478d removed examples/agents/deep-agent-supervisor. examples/agents/kitchen-sink does not register agents.

What's still owed

Review feedback on agents-wg PR 22, revised against bfc43abb and surfaced here for approval before posting (draft in mcpcontrib proposals/agents-wg/pr22-comments-draft.md):

1. Pre-merge identifier (draft A.1). Ask what identifier implementations should use for cross-implementation testing before the SEP merges, given mcpkit ships io.mcpkit/agents to avoid squatting the final name.
2. Stateless datapoint (draft A.3). SEP-2575 parity came for free because only discovery is new wire surface (TestStatelessWireParity).
3. Step-up questions (draft C, trimmed). bfc43abb adds a MAY for step-up on known-but-unauthorized agents. Two questions remain open: how agent-level gating relates to per-tool authorization, and whether the roster itself should hide agents the caller cannot use.
4. delegateTool (draft B): no longer a question. agents-wg PR 22 rejected the single-tool shape, so mcpkit drops delegateTool (tracked in [tracker] experimental/ext/agents: align wire shape with the agents-wg agent discovery SEP draft #1393). Mention it only as a datapoint that a second implementation tried it and is converging.
5. What to point at: experimental/ext/agents and its Go client decoder (clients/go/). No runnable host is claimed.

Done when

The remaining comments are posted on agents-wg PR 22 and the thread has been checked once for replies.