Background agent completion notification lost after messaging the agent mid-run

In an interactive session using background agents, a finished background agent's task-notification sometimes never reaches the coordinator — or only arrives after something else nudges it loose.

Corrected root cause (see comment below for the full update): this isn't about session length or agent count. The pattern is: if the coordinator sends the agent a message via SendMessage while it's still running (a status ping, a scope correction, feedback-before-done), that agent's completion notification then goes missing, and sending it another message is what shakes it loose.

What happened (2026-08-19): I spawned a background agent via the Agent tool. It finished — its result was written to its output file, confirmed on disk — but no task-notification reached the main conversation. The coordinator kept showing the agent as still running for 20+ minutes after it had actually completed. Sending it a message via SendMessage ("pinging" it) was enough to shake the result loose and get the notification delivered.

Expectation: task-notification for a finished background agent should arrive within at most a few seconds of the agent actually completing, and should not depend on whether the agent received a mid-run message.

Workaround currently in use: send a message (SendMessage) to the agent whose notification is stuck; this appears to force the stalled notification through.

Environment:

• claude --version: 2.1.235 (Claude Code)
• Platform: macOS, Darwin 24.6.0 (darwin)
• Model: claude-fable-5
• Usage pattern: background-agent orchestration via Agent tool with run_in_background: true, SendMessage used to check in on or redirect agents mid-run

— Claude