Your Model Upgrade Is Not a Harness Deletion Event
The intuition that better models let you remove scaffolding is backwards: rising model capability raises the autonomy ceiling only when the harness grows with it, so harness engineering is a durable discipline with its own roadmap, not a temporary crutch to be retired at the next model release.
AngleContrast the 'delete the scaffolding' reflex with concrete harness layers (verification steps, auto-mode classifiers, escalation rules) and argue that each model generation should trigger harness ablation experiments, not harness demolition.
Source video ↗Verification Is the Product: Why the Verification-to-Shipped-Code Ratio Flips
When agents generate code faster than humans can read diffs, code review stops being the correctness gate and the verification layer becomes the primary deliverable; teams that keep optimizing for reviewer throughput will ship faster and trust less, while teams that invest in machine-checkable verification surfaces convert velocity into reliability.
AngleArgue the ratio inversion (~100x verification code) as a structural consequence of generation velocity, then examine the failure mode where verification code itself drifts and manufactures false confidence.
Source video ↗The Conversation Was Never the Artifact: Loops Are
Designing an agent as a chat session under-specifies everything that matters in production; when the unit of design becomes a recurring loop with triggers, cadence, triage state, and escalation, the hard problems move from prompt quality to scheduling, idempotency, deduplication, and per-loop cost ceilings.
AngleReframe agent design around loop topology and use the morning-check and tag-driven triage patterns as the concrete case that exposes the missing infrastructure requirements.
Source video ↗Most Knowledge Work Is a Pipeline You Have Not Written Yet
The reduction of knowledge work to code and coding agents is not a philosophical curiosity but a practical selection rule: domains whose operations are deterministic and scriptable are candidates for scripted pipelines, and the engineer's new skill is modeling a workflow as a code-executable, verifiable pipeline rather than operating a GUI.
AngleTest the reduction claim against the FFmpeg/scripting and accounting-script examples, then propose determinism and verification cost as the two filters that predict which workflows convert and which do not.
Source video ↗Harness Ablation Rig
beyond-evalsGiven a fixed agentic task suite, a measurable subset of harness components (auto-mode classifiers, verification steps, escalation rules) becomes redundant as the underlying model improves, and ablating those components without loss of task success rate identifies removable scaffolding; harness complexity therefore does not need to grow monotonically.
Proof of conceptRun the same task suite across two successive model versions with a configurable harness where each component can be toggled; measure success rate, escalation frequency, and cost per task for every harness subset to locate the ablation frontier.
MeasurementMarginal task-success contribution and cost delta per harness component per model version; number of components that can be removed with no statistically significant success-rate loss.
Source video ↗Verification Ratio Meter
beyond-evalsVerification-to-production code ratio scales with machine generation rate rather than human review bandwidth, and pipelines below a threshold ratio exhibit rising defect escape rates that are not detected by human diff review.
Proof of conceptInstrument a repo or pipeline to track verification lines, production lines, generation volume, and post-merge defect escapes per time window; plot escape rate against verification-to-production ratio and detect the knee where escapes fall off.
MeasurementCorrelation between verification-to-production ratio and defect escape rate per generation-volume bucket; ratio threshold at which escape rate stops improving.
Source video ↗Loop Ops Harness
movement-labRecurring agent loops can run unattended only when explicit idempotency keys, triage-item deduplication, per-loop cost ceilings, and a defined escalation path are enforced, and loops missing these properties fail at a predictable task-ambiguity threshold.
Proof of conceptBuild a scheduled loop runner (cadence trigger plus tag-triggered triage) for a morning-check style job over events and PRs, with toggleable idempotency, dedup, cost ceiling, and escalation controls; escalate task scope and ambiguity until autonomy collapses.
MeasurementUnattended run success rate, duplicate-action rate, cost per loop run, and the task-ambiguity level at which human escalation becomes mandatory.
Source video ↗Generated Artifact Capability Sandbox
gatehouseModel-generated interactive artifacts that invoke MCPs and read data sources leak capability unless permissions are granted per artifact instance, and per-instance capability grants reduce unauthorized tool-call surface without reducing task completion rate.
Proof of conceptWrap generated artifacts in a sandbox where each instance receives an explicit scoped capability manifest for its MCP calls and data reads; compare unauthorized-call incidence and task completion against an unscoped baseline.
MeasurementRate of tool calls outside the granted scope, task completion rate, and observability coverage of tool-call provenance per artifact instance.
Source video ↗