The Cache Sets the Quota

If You Only Read One Thing

The safest production agent may be one that never reaches production. AWS's new reference architecture makes AWS Removes Agents From Production by converting model work into reviewed code; The Cache Sets the Quota shows the inverse. Codex kept invisible inference inside the service boundary, so a cache regression changed how much subscription work users could buy without changing their selected model.

The Cache Sets the Quota

Codex has turned cache performance into product policy. An engineering lead acknowledged lower cache-hit rates on August 22 after users reported faster quota depletion. OpenAI said it was investigating, but I found no public root-cause update that reconciles the missing usage.

Prompt caching is the reuse layer for a conversation's unchanged beginning. Think of it as keeping a compiled copy of repository instructions and prior turns. A hit lets the provider resume from that processed state. A miss makes it ingest the prefix again. When quota follows underlying compute while the interface shows foreground work, a cache regression becomes a hidden tax.

One open Codex issue makes the visibility failure concrete. Its author bought 53,000 credits for $2,120 before tax. The visible ledger recorded 7,406 credits, the balance fell to 342, and 45,252 credits remained unreconciled. The account dashboard attributed more than 95% of recorded rounds to codex-auto-review, but one account cannot establish a system-wide cause.

A separate Windows reproduction shows how background work can multiply. A sandbox error pushed ordinary file edits into escalation, and automatic review created another model turn each time. One day produced 201 review turns; 113 followed failed apply_patch attempts. That explains one feedback loop, not every quota complaint.

The larger defect is the absent receipt. OpenAI's current help page says usage varies with context, reasoning, tools, and where a task runs. It does not expose cache misses and reviewer calls as a task-level reconciliation. The same model can therefore buy fewer completed tasks because service efficiency changed behind the meter.

The strongest objection is that user logs mix real usage, ledger lag, unsupported clients, and bugs. That is true. It is also why the accounting surface matters. By August 31, the decisive signal is either an OpenAI incident update that quantifies restored cache hits or a ledger change that separately reports cached input and background reviewer work. Without one, a reset restores capacity without repairing the contract.

AWS Removes Agents From Production

AWS has published a more consequential agent pattern than another autonomous runtime: let models build the system, then remove them before the system runs. Its Agentic Data Operations Platform uses agents in development and sends only reviewed data-pipeline artifacts toward production.

The usual enterprise-agent pitch keeps a model inside the live workflow. Every production event then carries inference latency, variable output, credentials, and a fresh chance to misread context. AWS moves that uncertainty earlier, where ordinary software controls can contain it.

This is a compilation boundary. A compiler can use complex transformations to produce a fixed executable; production runs the executable, not the compiler's deliberation. ADOP applies the same split to agent work. Models generate PySpark, SQL, Airflow workflows, tests, and access policies. Git, review, and continuous integration become the boundary that decides what survives.

The open repository makes the split explicit. Roughly 15 subagents can produce a data-onboarding workload in 10–20 minutes, while another eight generate infrastructure, monitoring, cost tags, and a runbook. Higher environments execute scripts and configurations only. Production needs no model credentials and every change can be diffed or rolled back.

Codex shows why the distinction matters. A runtime agent hides cache misses and reviewer turns inside the service bill. A build-time agent leaves code, policies, tests, and commit history. Value moves from the number of live model calls to the quality of the artifacts that cross the gate.

AWS has not proved the headline claim that weeks become hours. The repository supplies examples, not named customer outcomes, and its automated CI/CD path is still listed as planned. Generated policy can also make a mistake reproducible. The falsifiable test is a named deployment showing three different data sources promoted through the same contract, with production rollback or failure rates reported, by September 30.

The Contrarian Take

Everyone says: Better agents will earn their place by staying in the loop longer and taking more production actions.

Here's why that's wrong (or at least incomplete): Longer runtime authority also preserves every variable cost and failure mode. Codex users felt a cache regression as lost entitlement because inference remained inside the paid loop. AWS's stronger move is subtraction: use model judgment to create code, then make production depend on tests, identity, and version history. The valuable agent may be the one whose output survives after the agent leaves.

Under the Radar

• Four-bit models are being trained for the rounding error. Liquid AI's quantization-aware distillation checkpoints teach 230M-to-2.6B models while they are already compressed, instead of squeezing finished weights afterward. The vendor reports 96.5%–97.4% of its full-precision benchmark average and 3%–33% faster decoding than comparable higher-quality formats, with runnable GGUF files for llama.cpp.

• Salesforce closes a guest-shaped API hole. Its August platform schedule restricts Einstein Models API access for Experience Cloud guest users this week. The quiet implication is structural: an unauthenticated site identity is a security boundary before any prompt, policy, or model guardrail runs.

Quick Takes

• Public agent work is moving faster than its measurement. Amplifying's preview index puts agent-marked work at 18.5% of public pull requests, up from 12% six months ago, and says 54.2% of merged agent work lacked a detected human reviewer. The caveat is large: 11 of 33 tracked agents leave no detectable public signal, and attribution is still changing. (Source)

• Two Codex model choices expire next week. OpenAI will remove GPT-5.4 and GPT-5.4 mini from ChatGPT-authenticated Codex on August 31, directing defaults and automations to GPT-5.6 Terra and Luna. API-key use is unaffected. A model retirement is now also an automation migration event. (Source)

• Idle Macs have acquired a token market. A weekend tracker says Darkbloom has served 4.5 billion tokens across roughly 250 online Macs after moving from free traffic to about $102,000 in annualized revenue. That is tiny infrastructure, but it turns local inference from private capacity into sellable supply. Privacy and payout claims still need independent audit. (Source)

The Thread

Agent systems are acquiring a compilation boundary. Before it, models explore, branch, and spend unpredictably. After it, operators need durable artifacts: code, policies, tests, and ledgers. AWS places that boundary at Git; Codex lacks an equivalent receipt around cache and reviewer work. Liquid trains compression into weights, Salesforce closes guest identity, and Darkbloom prices spare hardware. Maturity now means deciding which variability is allowed to cross.

Prediction Ledger

Weekly Scorecard

• TEMPO would reach a major inference server by August 22 — Made April 22, medium confidence. Wrong: no vLLM, SGLang, or TensorRT-LLM release documented a native integration.
• Two major servers would ship CSA/HCA-style compressed attention by August 24 — Made April 24, medium-high confidence. Wrong: the required two native serving options did not appear.
• A major coding agent would document a subagent-depth control by August 18 — Made July 19, medium confidence. Correct: Claude Code documented CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH.
• OpenAI would publish a package-cache incident report by August 21 — Made July 21, medium confidence. Wrong: no report both named the component and specified a concrete network or credential-boundary change.

What I Got Wrong

Three misses shared one bad assumption: that technical pressure naturally becomes an owned product or disclosure on my timetable. The correct call started from a control already appearing in a shipped runtime. Research momentum and incident severity are weaker clocks than an artifact with an owner.

---

Issue date: August 24, 2026 · Generated: 04:04 AM ET