Summary

Add non-blocking background agent execution as agent-callable tools. Unlike delegate_task which blocks the parent until all children complete, this lets the parent agent spawn long-running background agents and continue working immediately.

Motivation

delegate_task is synchronous — the parent agent is blocked in a tool call waiting for all subagents. This prevents Hermes from implementing workflows like:

• Start a long research/coding agent and do other work while it runs
• Run multiple independent agents across turns without waiting
• Check in on running agents periodically and steer them
• Collect results only when convenient

Proposed tools (new async_delegation toolset)

Example workflow

# Start three research tasks
t1 = delegate_task_async(goal="Research CDAL environmental violations")
t2 = delegate_task_async(goal="Find relevant ECSC case law on nuisance")
t3 = delegate_task_async(goal="Draft outline for affidavit")

# Check progress while doing other work
check_task(t1["task_id"])  # → {status: running, output_preview: "..."}

# Steer one mid-run
steer_task(t2["task_id"], "Focus on post-2020 cases only")

# Collect when ready
results = [collect_task(t["task_id"]) for t in [t1, t2, t3]]

Design

• In-process threads — reuses the existing AIAgent machinery, same credentials/toolsets as delegate_task
• Output capture — child's _print_fn redirected to an in-memory ring buffer (200 lines × 500 chars)
• Steering — direct _steering_injection queue wiring, same as synchronous subagents
• Registry — parent_agent._async_tasks dict keyed by task_id, session-scoped

Relationship to #4949

This is intentionally different from #4949 (Persistent ACP background subagents):

• Persistent ACP background subagents #4949 proposes ACP-over-stdio with sandboxed containers and cross-turn persistence — the right long-term architecture for production multi-agent systems
• This PR implements in-process threads with no external dependencies — works today, useful for single-session workflows

These are complementary. When #4949 ships, the API could be unified under the same tool names.

Tests

68 unit tests covering all 6 tools and the AsyncTask dataclass. No real LLM calls — all child agents are mocked.