Writing

The Real Reason Sub-Agents Exist (It's Not Parallelism)

April 11, 2026·
AIagentsarchitecture

Everyone talks about sub-agents as a way to run tasks in parallel. That's not the point.

The point is keeping your context clean.

I've been reading Dan Farrelly's breakdown of production agent systems, and the insight that stuck with me is simple: a sub-agent isn't a specialist. It's a compression boundary. The parent agent never sees the full execution trace — 15 tool calls, 8 files read, hundreds of intermediate steps. It only sees the summary. That's 90%+ fewer tokens added to the parent's context per delegation.

This matters more than it sounds. You can have a 200K token window and still get degraded outputs — not because you ran out of space, but because you filled it with noise. Models lose focus in bloated contexts. The quality drops before the limit hits.

Three Modes, One Principle

There are three ways a parent agent can delegate: synchronously (block and wait for the result), asynchronously (fire and forget, the sub-agent delivers directly to the user), and scheduled (run this later, with real-time data at execution time).

Sub-agent patterns: sync, async, scheduled

The right choice depends on whether the parent needs the result to continue. If it doesn't, async is almost always better — zero coordination overhead, free parallelism as a side effect.

The scheduled mode is the one people sleep on. It's not a cron job. It's an agent that runs at a future moment with the world's current state. "Check deployment metrics tomorrow at 9am" means actually pulling live data at 9am, not sending a pre-written message.

What This Means for Products with Real Users

I'm building a product where users come back repeatedly. Each returning session adds history: preferences, past interactions, context that matters. The naive approach is to keep stuffing it all into the prompt. That breaks fast.

The better approach: a sub-agent that reads the full history and returns one clean line to the parent. "This user prefers X, last came in Y weeks ago, sensitive to Z." That's all the parent needs.

Sub-agents aren't about doing more things at once. They're about making sure the things you do stay sharp.

← Back to Writing