OpenClaw Sub-Agents EXPLAINED (Stop Getting Slop From Your AI)

Sub-agents cut OpenClaw "slop" by splitting context and parallelising research — and the latest build now reports sub-agent failures instead of going silent.

Why a single agent gives you slop

Running research, writing, and graphics in one prompt forces the main agent to hold the entire context window and "rush to give you the results back." Boxmining spawns two research sub-agents in parallel — one reading OpenClaw docs, one searching the web — and the orchestrator reconciles the answers. If one sub-agent hallucinates, the other can correct it, and the main agent only needs to synthesise, not memorise.

Context window is the real reason this works

Sub-agents are deliberately "dumb agents" with a smaller context window. They don't need to know your morning routine or how you brush your teeth — only the orchestrator does. This is "context window optimization," and it matters even more on cheaper or Chinese models (MiniMax, Kimi, etc.) that "perform a lot better with smaller context." You can also configure sub-agents to run on different models than the orchestrator.

What's new in the latest OpenClaw

The previous version was "very wonky" — sub-agents would fail and "not even a notification that it failed." The current build gives you an update when a sub-agent fails, so you can re-run it instead of getting an empty response.

Prompt pattern that actually triggers parallelism

By default, OpenClaw agents do not run parallel sub-agent workflows unless you ask. A working prompt template: can you make a presentation on sub agents for openclaw? Go send sub agents to research why it's important then come back to me. Use sub agent to make the presentation. And other sub agents to do the SVG graphics. Force the breakdown: "one agent do the graphics, one agent do the presentation."

The cost trade-off

Parallel sub-agents mean more total work and more tokens. On the MiniMax plan the prompts are bundled in, so cost is largely absorbed. On Opus, "maybe you should be a little bit worried" — the creator flags ~$30/day as the real Opus question, and their preliminary testing leans toward Opus for quality but not for price.

Model split the team is running

Stark is using MiniMax more, the host is using Opus more. They're publishing the Opus-vs-MiniMax comparison in a follow-up.