Feedback Loops

"Every assumption an agent makes without verification becomes a compounding error."

A linear pipeline looks clean on a diagram: input goes in, output comes out. But every node in that pipeline makes assumptions about the previous node's output — and never checks them. Toggle between the two views below to see what's actually happening between the boxes.

The signal matters more than the loop.

People conflate three fundamentally different mechanisms under "feedback loop": retrying on error, human review, and evaluator-as-judge. Most systems implement the first and claim they're doing the third. Click each tab to see what's actually different between them — trigger, signal, and what changes.

Generate · Evaluate · Delta · Steer — and the one node everyone skips.

A real feedback loop has four phases. Most implementations collapse two of them — they skip straight from Evaluate to Generate again, bypassing Delta entirely. Delta is the load-bearing node: it answers "what specifically changed, and why?" Without it, you're not iterating. You're sampling. Step through below to watch the context window build across passes.

An agent that doesn't know when to stop isn't autonomous — it's runaway.

Loops without convergence criteria are infinite regress. Most people solve this badly: stop after N iterations (arbitrary) or stop when it looks good enough (circular). The right approach is to define what "good enough" means before the loop starts — in measurable terms. Drag the quality bar to feel the cost/quality tradeoff directly.

Parallelism is spatial. Iteration is temporal. You need both.

Most teams treat parallelism and iteration as separate design decisions. They aren't. Parallel sub-agents — running multiple agents simultaneously to get diverse outputs — are what happens inside Generate. The feedback loop governs what happens across iterations. Spatial versus temporal. Click to zoom in and see where they connect.