Agent Team: critical usability problems with teammate communication and specialized agent types

Summary

The Agent Team feature has two significant implementation problems that make it unreliable in practice.

Problem 1: Specialized agent types (e.g. ben-sigelman, don-stewart) fail to start silently

When spawning a teammate with a specialized subagent_type (e.g. ben-sigelman, charity-majors, don-stewart), the agent is reported as successfully spawned but never sends an idle notification, never processes messages, and never produces output. It fails silently with no error surfaced to the team lead.

Reproduction:

• Create a team with TeamCreate
• Spawn an agent with subagent_type: "ben-sigelman" and run_in_background: true
• Agent is reported as spawned but sends no idle notifications and does not respond to SendMessage
• Retrying with a different name (e.g. ben-2, ben-sigelman) produces the same silent failure
• The agent appears to never actually start

In this session, ben-sigelman was attempted 3 times with different names and never started. The only way to unblock was to have the team lead run Ben's analysis inline.

Problem 2: Some agent types do not have the SendMessage tool

don-stewart (and possibly other specialized agent types) do not have SendMessage in their tool list. This means they cannot communicate with the team lead or other teammates at all. The agent is spawned, does its work, but has no way to report back.

Impact: The teammate goes idle repeatedly sending idle notifications (which the team lead sees as noise) but the actual output is silently written nowhere — or written to a file only if the agent independently decides to do so. The team lead has no way to know where to look.

Observed workaround: In this session, Don Stewart wrote his output to design/adr/ files and the team lead had to poll the filesystem to discover the output. This is not a documented or expected behavior.

Expected behavior

1. If a specialized agent type fails to start, surface an error immediately rather than reporting success and going silent.
2. All agent types that can be spawned as teammates should have SendMessage available, or the framework should document which types do not and provide an alternative communication channel.
3. The team lead should be able to discover teammate output through a consistent mechanism, not by guessing filesystem paths.

Impact

These two bugs combined make the Agent Team feature unreliable for any multi-agent workflow that depends on specialist agent types. In the session that surfaced these bugs, 3 agent spawn attempts failed silently, a teammate had no way to communicate its output, and the team lead had to manually perform the work of the failing agents inline.