Sub-agents that inherit the gate

Give the agent a job with several independent parts — research three markets, draft five variants, clean a pipeline while writing the weekly update — and it spins up a sub-agent for each, running side by side in isolated contexts. When they finish, the delegating agent wakes, integrates the results, and decides the next step.

Work is split so the parts do not overlap, and agents can claim shared resources before touching them — so two strands never silently overwrite the same thing.

• Independent work runs at the same time, not one after another
• Each sub-agent works in its own isolated context
• Shared resources are claimed before they are touched

A sub-agent has no special powers. It runs under the same tenant, the same tool policies, and the same default-deny gate as the agent that delegated to it: its mutation-capable calls become pending proposals in the same queue, and unknown tools fail closed for it exactly as they do at the top level.

That inheritance is the difference between delegation and dilution. Parallelism multiplies throughput; it does not multiply the set of things that can happen without you.

• Same policies, same proposal queue, at every depth
• No privilege escalation through delegation
• Unknown tools fail closed for sub-agents too

You stay in command while the fan-out is live. A message folds new instructions into work already in flight, and a stop command halts an agent and everything it spawned, immediately — no orphaned sub-agent keeps running in the background after you have said stop.

Interrupted work is preserved, not discarded: the delegating agent can read partial progress, judge it, and continue from where things stopped rather than from zero.

Every delegated task carries a definition of done. The sub-agent checks its own output against that bar, and the delegating agent independently validates the result before accepting it — work that misses the goal is sent back with the reason, not waved through. And every step each sub-agent took is on the record.

The trace matters as much as the check: you can reconstruct exactly how any piece of parallel work happened — which tools were called, what was produced, and what was approved along the way.