Your Retrieval Phase Should Emit a Skeleton, Not a Context Window
Most RAG and agent pipelines treat retrieval as context-stuffing for generation and therefore cannot localize failure or re-run a plan; making the explore phase terminate in a reviewable, machine-checkable structure artifact converts generation into schema-filling and turns prose failures into structure failures that are independently verifiable.
AngleReframe the retrieval/generation boundary as an interface contract, then show what becomes measurable (plan diffs, structural coverage) and what becomes cheap (re-running the plan) once that contract exists.
Source video ↗Kill It at Ten Percent: The Cheap Gate That Should End Most Agent Projects
Abandonment at a scaled proxy gate should be an expected outcome rather than a failure signal, because a production-shaped artifact at roughly a tenth of full cost collapses the opportunity cost of dead ideas — but the gate only earns its keep if proxy size is swept empirically instead of inherited from folklore ratios.
AngleTreat the proposal-to-book ratio as a design parameter to be measured, define the false-positive rate it must control, and argue that teams without a cheap kill gate are really paying full price for their own ambiguity.
Source video ↗Stalled Pipelines Don't Need More Data, They Need a Different Axis
When throughput collapses in a long-horizon system, the default reflexes — add retrieval volume, scale the model — are the expensive and often wrong fixes; re-indexing or re-decomposing along a different axis is a cheaper first move that can dominate data volume as a cause of stalled progress.
AnglePresent structure/index choice as a first-class tunable with a cheap A/B: same corpus, two orderings, compare output-rate change before and after the switch.
Source video ↗Interruptions Are a Scheduler Problem, Not a Prompt Problem
Long-horizon output is sustained by an interference policy plus a declared per-window throughput quota, not by motivational prompting; teams that leave inbound async events unthrottled are choosing burst-and-crash throughput over predictable variance and should own that choice explicitly in the scheduler.
AngleMap email/pings to async inbound interrupts in agent runs and argue for exclusive quiescent windows with a units-per-window budget, accepting added latency for interrupt-driven work as the stated price.
Source video ↗