AI Delivery Is Now a Loop Problem

Two engineering posts landed two days apart — Hightouch on November 4, Anthropic on November 6 — arguing the same thesis from opposite ends of the stack. Anthropic published a breakdown of the memory it gives Claude across sessions S1. Hightouch reported benchmark experiments testing whether a cheap decision model, escalated to a frontier model only when unsure, can grade identity resolution S2. Neither announces a new model; both are about the loops around models — what gets remembered, corrected, and spent. That is increasingly the work delivery teams are hired to do.

Salience beats storage

Anthropic compared memory folders across Claude models playing Pokémon Red. Sonnet 3.7 archived everything: 413 files across 225 folders, one badge earned; Fable 5 wrote nine files in two folders plus a dashboard note and earned four badges. The archivist burned effort refreshing facts the screen already showed; the editor kept strategy and rules that vanish unless written down. Steering matters more: telling Opus-class models to "record only failures" cost much of memory's benefit; leaving salience judgment to the model worked better. The injected layer stays small, too; Claude Code loads only the first 200 lines or 25 KB of its MEMORY.md index. Ship a memory policy, not a rulebook.

Correct the record offline

Notes can be wrong from birth. One run had Claude misremember which side of a mountain a Pokémon Center was on, trust the faulty note onto a one-way ledge three times, and never reach the next town. A "dreaming" pass that replays transcripts against stored notes caught and rewrote the error; the corrected run went straight through. At corpus scale, agents building a linked wiki over roughly 350,000 internal documents (~420M tokens) lifted a 25-question evaluation from 13% to 48% of graded criteria as dreaming scaled from two to 128 sessions. Wiki plus search beat search alone, 61% versus 53%, and saved about 20,000 tokens per question. The operational read: agent memory needs scheduled maintenance, like a search index. Anthropic also watched a wiki outgrow the documents it summarized, and someone must own the pruning.

Escalate on doubt, not by default

Hightouch's evidence is a benchmark, not a deployment: identities synthesized with pseudopeople, a University of Washington tool simulating a fictional US town — 1,000 identities, half clean, half deliberate over-merges. Grading those, Jev alone scored 95.6% at $0.05 per 1,000 identities; routing the 30–90% confidence band to Claude Opus 5.5 reached 98.8% at $0.88, versus 99.0% at $3.86 for Opus on everything. On a harder benchmark the hybrid edged Opus alone, 95.3% to 95.1%, at just over half the cost, as Jev's calibration pushed more cases upstream when difficulty rose. GPT-6 Luna looked stronger out of the box at 97.3%, but its overconfidence capped its hybrid at 91.6% on the hard set. Calibration, not raw accuracy, makes escalation pay — promising results, with a production track record still to be earned.

The deployment job description

Neither post is a production case study — Anthropic reports internal evaluations, Hightouch synthetic benchmarks — but together they describe the loops someone must own once a system ships: a salience policy for what agents write down, a maintenance pass that checks notes against ground truth, and a spend policy that routes on calibrated uncertainty. None live on a model card; they live in deployment engineering. Y Combinator's founding-engineer interviews describe the same shape on the human side — engineers who "own different verticals" well beyond code S3. The forward-deployed engineer's artifact in 2026 is not a prompt; it is the loop that keeps judgment honest after the demo ends.