Your Context Window Is Not a Bucket, It's a Room With Bad Acoustics
Context capacity is the wrong optimization target: because attention is positionally uneven, teams should engineer placement, ordering, and segmentation of instructions rather than maximize token budget — and a longer prompt is a measurable regression, not a neutral trade.
AngleReframe long-context work as a placement problem, with concrete harness consequences: put invariants at the edges, prune the middle, segment rather than concatenate, and test which regions the model actually attends to before trusting a spec file.
Source video ↗Modularity Was Always About Context — AI Just Made the Bill Visible
Module boundaries and dependency-injection seams should now be justified partly by how much context they isolate; a leaky interface is no longer only a human-comprehension problem, it is a direct cause of model confusion, which makes refactoring toward cohesion an AI-reliability intervention.
AngleConvert classic cohesion/coupling arguments into context-budget metrics and show how poor boundaries let unrelated concerns bleed into the agent's working set, causing conflation that looks like a model failure but is an architecture failure.
Source video ↗Stop Measuring Agents in Lines Generated: A Cleanup-Cost Accounting
Agent productivity claims are inflated because they exclude rework; net cycle time including cleanup is the honest metric, and it implies agents pay off first in codebases that already have cheap cleanup — tests, coverage, mutation testing, clean boundaries — rather than uniformly across all projects.
AngleIntroduce a rework-inclusive accounting model, then argue that the discipline infrastructure skeptics dismiss as ceremony is precisely what gates the return on agent adoption.
Source video ↗Tactical Is Automatable, Strategic Is Not: A Division of Labor for Agentic Development
Agents are reliable at making something work now and unreliable at long-horizon design, so the human role should be reallocated to boundaries, constraints, and invariants — and design-level delegation, not code generation, is where agentic projects actually drift.
AngleSeparate meta-level decisions from execution-level decisions, define the failure signature of misplaced delegation (design drift, wandering off track), and specify what humans must still own even when they stop writing code.
Source video ↗Attention Atlas
movement-labInstruction-following and recall accuracy in long prompts is a measurable, non-monotonic function of token position, and critical constraints placed mid-context fail at a rate detectably higher than identical constraints placed at the beginning or end.
Proof of conceptBuild a probe harness that places identical instructions and needles at N controlled depths within a synthetic long context, sweeps total context length across several buckets, and records per-position compliance and recall for multiple models and task types. Emit a position-vs-accuracy curve per model plus a recommended placement policy for harness assembly.
MeasurementPer-position accuracy curves; the depth and context-length at which mid-context accuracy falls below edge accuracy by a fixed margin; regression of total context length on overall task success.
Source video ↗Boundary Benchmark
gatehouseFor identical feature tasks, agents given module-scoped context achieve higher task success and lower token consumption than agents given whole-repository context, and the gap widens as repository size grows.
Proof of conceptTake a real repository, produce two variants — an original low-cohesion structure and a refactored high-cohesion one with disciplined interfaces — and run the same task suite against both under fixed model, prompt, and tooling, logging success, tokens, retries, and human-judged rework per run.
MeasurementTask success rate delta, mean tokens per completed task, retry count, and rework minutes per variant; correlation of the delta with repository size and module cohesion metrics.
Source video ↗Pipeline Ablation Bench
beyond-evalsEach added pipeline stage (cleaner, hardener, reviewer, QA) yields diminishing marginal defect reduction, and beyond a threshold stage count the coordination overhead — handoff loss, latency, token cost — exceeds the quality gain.
Proof of conceptImplement a configurable pipeline with independently toggleable specifier, coder, cleaner, hardener, reviewer, and QA roles over a seeded benchmark of tasks with known injected defects; run full stage ablation, recording defect escape rate, mutation score, latency, and cost per completed change.
MeasurementMarginal defect-escape reduction and mutation-score gain per added stage; cost and latency per stage; the stage count at which marginal quality gain falls below marginal coordination cost.
Source video ↗Gate Parity Study
gatehouseFor agent-authored changes, a mutation-score gate prevents more escaped defects per unit cost than a line-coverage gate, and the two combined add less than their individual contributions when human review is already present.
Proof of conceptRun agent-authored changes through three gating regimes — coverage-only, mutation-score-only, and combined — both with and without a human review step, then track defects that escape to the mainline plus the cost of each gate.
MeasurementEscaped-defect rate per gate regime, defects caught per gate-minute and per gate-token, marginal contribution of mutation testing over coverage, and interaction effect with human review.
Source video ↗