2026-08-20

1. OpenAI says frontier monitoring doesn't need your data. Anthropic has been requiring it since June — and voiding zero-retention contracts to get it

• OpenAI, 2026-08-19: Offering Zero Data Retention for frontier models restates ZDR for eligible API customers — nothing retained after processing, no personnel review, no training use without opt-in, customer-held encryption keys — and previews Private Safety Processing, said to detect misuse patterns across related interactions while sending OpenAI only a narrowly defined safety signal, without exposing the underlying prompts or responses. Enterprise and API only; rollout and a technical white paper in September (source).
• Anthropic, since 2026-06-09 ⚡ ** traffic is retained 30 days, on first- and third-party surfaces. It overrides a negotiated zero-retention agreement with no opt-out, and Fable 5 does not support ZDR at all (source).
• The comparison is not yet like-for-like: one policy is enforced and 72 days old, the other is a preview with a promised paper.
• Why it matters: both labs accept the same premise and only one can be right about whether it forces content retention — the answer decides whether "zero retention" survives as a procurement category for frontier models. And neither has published a detection accuracy or false-positive rate, the third safety-relevant classifier in two weeks announced here without one.
• → Safety Monitoring and Data Retention, Claude Fable 5, OpenAI, Anthropic

2. The harness stopped being a deployment choice — it is now inside the weights

• Agent Lightning v1.0: Towards Harnessed Agentic RL (arXiv:2608.17528) names the regime: harnessed agentic RL, where the deploy-time harness owns the environment interaction loop and the trainer sees only LLM request/response pairs through an endpoint proxy. Four other frameworks are stated to have adopted it. Reported: Qwen3.5-9B 41.8% → 56.4% on SWE-bench Verified from 6K examples.
• This is the mechanism behind ClawGym II: Exploring Black-Box RL on Agent Harness (arXiv:2608.16798)'s finding yesterday that a harness-trained model stops being harness-agnostic.
• The competing bet landed the same day: Agentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Requirements (arXiv:2608.17310) uses evolution strategies to delete the credit-assignment machinery rather than fix it — +6.69% over a No-Skill baseline on WebArena-Lite, matched baseline beaten in 28 of 36 settings. Not comparable to the above: different model, benchmark, baseline and convention.
• Why it matters: this wiki's rule is that a benchmark number is a claim about a (model, harness) pair. If this is the default post-training route, that stops being a reporting convention and becomes a fact about the weights — a harness-agnostic model is something to demonstrate, not assume. The paper arguing hardest for the harness never names the one behind its own 56.4%.
• → Agentic Reinforcement Learning, Eval Harness Configuration

3. What a skill actually does, measured — and it is not what the word suggests

• Demystifying Agent Skills: Why They Work — Until They Don't (arXiv:2608.14036) open-codes 8,135 trial records: procedural anchoring accounts for 65.7% of skill cases against 4.5% for explicit knowledge injection. A skill library is a runbook, not a knowledge base. Skills beat Workflow Memory by +6.06 points.
• The separate failure is the day's sharpest number: actual-use precision falls 29.6% → 3.3% as the pool grows from 5 to 100.
• Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents (arXiv:2608.15008) reports the same shape from the memory side — one harness, 26 metrics, 3 backbones, 4 suites, and no substrate dominates: broad retrieval helps long-context factual QA and harms sequential decision-making by pulling attention off action-critical context.
• Why it matters: this is the mechanism under the reading adopted here yesterday — harness scaling buys execution reliability, not self-assessment — reached by a third method. And it says retrieval quantity is not monotone in usefulness: every skill and memory figure this wiki holds is reported at one small pool size, on one task class, and both now look load-bearing.
• → Agents (LLM Agents), Eval Harness Configuration

4. Frontier open weights became runnable on one machine, in the week the memory to run them repriced 5×

• FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution (arXiv:2608.16157) serves 20+ MoE models from an 8GB laptop GPU upward — 35B on a laptop, 284B on a gaming desktop, and 753B GLM-5.2 on a single workstation GPU — by remapping computation onto free resources instead of fixing an offloading strategy.
• Against it: consumer DDR5 up as much as 485% year over year; a 128GB kit at $3,399 against a lowest tracked price near $329; 2×32GB kits $222 → $1,272. No forecast read expects relief before late 2027 (source).
• Why it matters: FreeToken's method substitutes host memory and bandwidth for GPU capacity — it spends precisely the resource that just repriced. Also worth holding: every FreeToken figure is capacity, none is throughput, and no quantisation level or accuracy check is published. Fitting a model is not serving it.
• → Open-Weights Policy Fight, NVIDIA