Named Agent tool spawns (name: param) accepted but never executed — no transcript, no TaskList entry, no notification

Environment

• Claude Code CLI on Linux (tmux-managed long-running sessions)
• Reproduced across two independent projects/panes on 2026-07-19 and 2026-07-27,

and again 2026-07-29

• Models: reproduced with both sonnet and haiku subagent models

Summary

When the Agent tool is invoked WITH a name: parameter, the spawn is accepted —
the tool result says the agent "is now running and will receive instructions via
mailbox" and returns an agent_id of the form name@session-… — but the agent
never executes: no transcript file is created, no TaskList entry appears, no
completion notification ever fires, and SendMessage wake probes to its inbox get
no response. Unnamed spawns issued minutes apart in the same session run
normally.

Repro (controlled, 2026-07-27)

1. In a long-running session, call Agent with `{name: "probe1", model: "haiku",

prompt: "echo hello and finish"}` → acceptance: "agent_id: probe1@session-…,
will receive instructions via mailbox".

1. Wait 25+ minutes. Observe: no TaskList entry at all for the named agent, no

output file, no notification.

1. Send SendMessage {to: "probe1", message: "status?"} → delivered to its

inbox, never answered — consistent with the agent process never having
started, not with a notification-path failure.

1. Control: two unnamed async Agent spawns in the same window (same models,

comparable prompts) both ran and completed normally.

Impact

Named spawns are the documented way to keep an addressable handle on a subagent
(SendMessage continuation). Silent non-execution means orchestration built on
named agents stalls indefinitely with no error surface — we lost 9.5h to a
stalled 4-agent fan-out before diagnosing (all four named spawns accepted and
dead; four unnamed retries ran fine).

Notes

• The acceptance shape differs between the two paths: named returns "instructions

via mailbox", unnamed returns "async launched + output_file" — the break looks
like the initial mailbox prompt delivery on the named path.

• A later named spawn on 2026-07-29 DID start (its task could be TaskStop'd), so

the failure may be intermittent or state-dependent (long-lived session,
many prior spawns).

• Fourth occurrence, third independent project (2026-07-29): a batch of named

research subagents ("dr-*") spawned for a scheduled daily workflow never
reported back; the parent session had to fall back to doing the synthesis
itself. Live-workflow impact, not just orchestration tests.

Happy to provide session IDs / timestamps privately if useful.