Aaron Ahmad

My LLM's LLM

17 August 2026 · tech
Patent-style schematic of nested LLM delegation: conductor directing a specialist

I used to drive the specialist coding model directly: paste the problem, wait, paste the next. It worked, but I was the router.

Now I direct my main agent, and it decides when a problem is hard enough to hand to the specialist. The specialist grinds it headless and hands back a result. The main agent QA's it before I see it. My LLM has its own LLM.

FIG. 1 SHEET 1/2 THE CHAIN SCALE N/T DELEGATION USER OPERATOR DIRECT HERMES AGENT CONDUCTOR · HOLDS CONTEXT CONTEXT HISTORY · INTENT DELEGATE ISOLATED WORKSPACE THE ORACLE SPECIALIST · HEADLESS GRIND RESULT QA CHECK VERIFIED RESULT LEGEND: COMMAND FLOW CONTEXT LINK ISOLATION BOUND.
FIG. 1: The chain. Operator to conductor to specialist, QA, and back.

The benefit is separation of concerns. The main agent holds the thread: history, intent, the shape of the day. The specialist holds only the problem in front of it. I talk to one interface, not two.

FIG. 2 SHEET 2/2 THE LOOP SCALE N/T QA CYCLE DELEGATE, EXECUTE, QA, REDO STEP 1 STEP 2 STEP 3 STEP 4 USER REQUEST HERMES AGENT DELEGATE DELEGATE THE ORACLE EXECUTE HEADLESS OUTPUT QA GATE PASS USER DELIVER FAIL REDO · LOOP CAN REPEAT NOTE: THE SPECIALIST CANNOT ASK THE OPERATOR QUESTIONS MID-TASK. IT WORKS HEADLESS. THE CONDUCTOR MUST JUDGE WHETHER THE OUTPUT IS ACCEPTABLE OR MUST BE REDONE. THE LOOP CAN REPEAT INDEFINITELY. LEGEND: FLOW DIRECTION DECISION GATE REDO ARROW = FAIL PATH
FIG. 2: The loop. Delegate, execute, QA gate, pass or redo.

The trade-offs are real. The specialist can't ask questions mid-task. Delegation adds latency. The main agent can bottleneck, and I must trust its judgement on when to delegate and whether the result is good. Cost: a flat subscription for the everyday model (~£15/mo) plus usage-based credits for the specialist. Ballpark £15 to £60/mo.

The bottleneck moved from my time to the delegation layer's judgement. Worth it.

← All posts Reply via email