Technically speaking

A planner agent breaks your request into tasks, then assigns each task to a specialist agent with a clear role.

• It is like a pit crew, each person has one job so the whole team moves faster and cleaner.
• Example: The planner creates a small mission list, delegates in parallel, then combines everything into one final answer.

Sub-agents run concurrently, each focused on one objective. The system waits for all outcomes, then combines them.

• Like cooking with three burners instead of one, dinner is ready faster with less back-and-forth.
• Example: Three independent tasks launch together, then the final answer is assembled from all three outputs.

lmkgpt exposes an API endpoint so external systems can trigger the same multi-agent workflow.

• It is like giving your software a hotline to the same research squad you use in the UI.
• Example: A product can submit a prompt server-to-server and receive structured outputs plus the final summary.

Progress is streamed from server to page continuously, so updates appear as agents produce results.

• The app keeps a live line open and sends small updates as they happen.
• Example: Agent statuses and text stream in increments, not as a single delayed payload.

The app keeps the planning steps, per-agent objectives, and raw outputs visible for review.

• It is like seeing the rough draft, notes, and final essay, all in one place.
• Example: You can verify that the final answer matches what the agents actually produced.

The system prefers one provider first, then falls back when errors are retryable.

• It keeps an alternate route ready so one traffic jam does not cancel the trip.
• Example: Each agent can recover independently, so one timeout does not collapse the whole run.