Agent teams: per-spawn teammate display mode + per-teammate MCP scoping (lean boot for large fleets)

Context

We run agent teams heavily (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, CC v2.1.204/205) — both small watchable judgment teams (3–5 workers we steer live in panes) and large mechanical fleets (20–30 workers doing batch formatting/verification, nobody watches them individually). Two session-level behaviors that are per-team concerns in practice:

Ask 1: per-spawn teammate display mode

teammateMode (and --teammate-mode) is locked at session launch, but whether a team is watch-worthy is decided mid-session, per spawn. A recent run: a 30-worker mechanical fleet spawned under teammateMode: "auto" opened 30 iTerm2 panes; TaskStop on wrap then raised 30 separate iTerm2 close-confirmation dialogs (each pane still had its MCP server child processes running), which the user had to dismiss one by one. The same machine also runs small teams where the panes are exactly what we want — the user reads a worker's stream and types course-corrections into its pane directly.

Request: a display override at spawn time — e.g. an Agent-tool parameter (display: "pane" | "headless") or a per-team option — so the lead can choose pane mode for watchable judgment teams and headless for mechanical fleets within one session. Today's workaround is choosing the mode before launch, which is the wrong decision point.

Ask 2: per-teammate MCP scoping (or a lean-boot option)

Teammates boot the session's full project+user MCP loadout. The mcpServers frontmatter scoping that works for custom subagents is documented as not applying to teammates, and disallowedTools restricts tool access but doesn't prevent the boot. For a 30-worker formatting fleet, that's ~30 redundant copies of every configured MCP server (web-search, scraping, cloud MCPs, etc.) spawned and never used — significant memory/startup cost, and on teardown those child processes are what trigger the terminal's close-confirmation friction in Ask 1.

Request: apply agent-definition mcpServers scoping to teammates too, or provide a spawn option like mcpServers: [] / --no-mcp for teammates, so mechanical workers can boot lean.

Environment

• Claude Code v2.1.204 / v2.1.205, macOS (iTerm2), CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
• Observed with teammateMode: "auto"; we've since defaulted to in-process, which resolves the dialog friction but gives up watchability entirely — hence the per-spawn ask.

Thanks — agent teams have been genuinely excellent for adversarial review and self-claiming work queues; these two knobs would make large fleets as ergonomic as small teams.