Your Database Was Built for Humans: Agent Query Loads Are a Different Design Target
Agentic query workloads are a distinct access pattern class — high-volume, non-interactive, differently concurrent — so human-optimized database assumptions about latency, indexing, caching and cost are not merely stressed but wrong, and teams should instrument agent workloads before they provision for them.
AngleTake each human-tuned design assumption and show the failure mode under agent-generated load, then describe what a replacement design target looks like.
Source video ↗Delegation Is a Capability Trade, Not a Productivity Multiplier
Handing cognition to an LLM buys throughput at a measurable cost to the delegate's underlying skill, and the same trade runs through delegating coding to agents — so the delegation boundary should be set by which skills must be retained, not by which tasks happen to be automatable.
AngleArgue the offloading finding generalizes from education to engineering practice, and propose a retention-first rule for deciding what humans keep doing.
Source video ↗Shaping the Build: Why the Scarce Skill Migrated One Level Up the Abstraction Ladder
As generative systems absorb execution, the scarce input moves from implementation to specification and selection, which means the tools and checkpoints that matter shift to agent orchestration, spec quality and verification of agent output — not authoring.
AngleTrace how the boundary of human work moves when execution is absorbed, and map the concrete tooling and review surfaces that follow.
Source video ↗A Skills Map Built From Job Postings Is a Map of Wishes, Not of Performance
A four-competency AI engineering taxonomy derived from 10,000 job descriptions and surveys captures stated requirements, so it should be validated against measured outcomes before it drives hiring, role design or upskilling budgets — especially the claim that the skills apply uniformly across roles.
AngleCritique the evidence basis and the uniformity assumption, then propose what an outcome-validated version would need to measure.
Source video ↗Agent-Load Database Fault Mapper
gatehouseAt least one specific design choice in human-tuned database engines (e.g. connection-pool sizing, cache eviction policy, interactive-latency indexing) degrades measurably faster under non-interactive agent-generated query loads than under an equivalently sized human interactive workload.
Proof of conceptReplay a synthetic agent workload (high-volume, non-interactive, bursty, retry-heavy) against a human-interactive baseline of equal total operations on the same engine, and instrument latency percentiles, pool saturation, index hit rates and cost per operation to locate the first breaking assumption.
MeasurementRatio of degradation in p99 latency, saturation onset time and cost per query between agent-mode and human-mode workloads; identification of which specific configuration knob or design assumption breaks first.
Source video ↗Delegation Boundary Experiment Harness
movement-labFor a fixed task class, the fraction of the task delegated to an agent determines near-term output but inversely predicts the human's retained ability to perform the same task unaided later — and a retained-practice split can hold output roughly flat while preserving capability.
Proof of conceptSplit engineers across three arms (full delegation, partial delegation with a mandatory human-practice component, no delegation) on matched tasks, measure immediate output, then re-test capability unaided at a delay using held-out tasks of the same class.
MeasurementImmediate output volume and quality versus delayed unaided performance on held-out tasks, per arm; effect size of the retained-practice split relative to full delegation.
Source video ↗Shaping-the-Build Signal Audit
beyond-evalsIn AI-augmented teams, hiring and promotion signals do not yet weight product intuition and architectural judgment above syntax proficiency, even though the stated work has shifted to steering coding agents and shaping the build.
Proof of conceptCode interview rubrics, promotion packets and hiring criteria for a sample of AI-augmented teams against how work is actually allocated between spec/judgment and implementation, exposing where signals still reward syntax recall.
MeasurementCorrelation between rubric/promotion weightings and measured allocation of work to spec and judgment versus implementation, and the size of the stated-versus-actual gap.
Source video ↗Four-Competency Outcome Validation
newThe four competencies (uncertainty management, SE fundamentals, agentic coding, shaping the build) predict measured engineering outcomes better than alternative skill taxonomies, but their cross-role overlap is not uniform — it is concentrated in specific role types rather than spread evenly across frontend, backend, DevOps and mobile.
Proof of conceptOperationalize the four competencies into observable rubrics, score engineers across roles against both these and a control taxonomy, then regress scores on measured delivery outcomes and defect/incident rates to test predictive power and role uniformity.
MeasurementComparitive predictive power (variance explained) of the four-competency model versus a control taxonomy, plus per-role competency distribution overlap and dispersion.
Source video ↗