Friday’s feed was agent-platform shaped: 10 keeps out of 945 gathered, and OpenAI shipped an Agents API the same day it put its full-duplex voice model in the API. The Agents API is the one to read closely — server-side sessions, subagent orchestration, and a self-hosted sandbox with a skills path, but US-only data residency and explicitly not ZDR-eligible, including when you bring your own sandbox. Around it: Cognition’s SWE-2 lands within a point of Fable 5.1 on its own benchmark at a claimed 64% lower cost, a 122B MoE trained to drive a real shell for 300+ turns, three agent papers including a prompt-injection detector that runs as a sidecar over existing logs, and OpenAI pausing new $200/month Pro signups while the API stays unaffected.

Lead — OpenAI ships an Agents API, plus GPT-Live-1 for voice

  • Agents API (product post) — Day-1 platform release and the most stack-relevant thing today: server-side sessions that persist across turns, multi-agent orchestration with max_concurrent_subagents, an environment block that can be self_hosted with a capability_directories path for skills, and web_search available in the same request. Two caveats to know before porting anything: data residency is US-only and the Agents API is explicitly not ZDR-eligible — including when you bring your own sandbox. Docs read in full; no pricing details on the overview page. (HN 272 · 150 comments)
  • GPT-Live-1 in the API — Same day: the full-duplex voice model that debuted in ChatGPT goes to the API. One model reasons over incoming and outgoing audio together (no STT→LLM→TTS chaining), handles interruptions mid-response, and can delegate reasoning and tool calls to a backend text model (Astra or a third party) while the conversation continues; telephony support for phone-call agents. The adoption numbers quoted in the post (Speak: ~80% fewer interruptions; a health company: 80% less code, 23K lines removed) are customer-supplied and quoted as such. (openai.com · Sep 10)

Agent frameworks & tooling

  • DriftNet: A Dual-Head Trajectory Transformer for Detecting and Localizing Prompt Injection in LLM Agents (arXiv 2609.10892) — Reads a logged tool-call trajectory and answers all three operator questions in one forward pass: is the trace compromised, where did the injection enter, which steps are hijacked. Under 2M parameters, frozen encoder, no access to the agent’s model — it’s a sidecar over your existing logs. On the task-disjoint split of AgentDrift (12,536 trajectories / 71,024 labeled steps, test evaluated once with a 20-config sweep bounding sensitivity to 0.011 F1): 0.983 trajectory F1, exact injection-point recovery on 98.7% of attacked traces, hijacked-span IoU 0.979, zero flags on 218 resisted attacks, 2.9% flags on hard negatives. Submitted Sep 9; abs verified. (arXiv cs.CR)
  • When Synthetic Data Hurts: On Catastrophic Forgetting in Skill Retrieval for LLM Agents (arXiv 2609.10750) — A production skill router over 34,396 skills, with the result that matters if you fine-tune anything on synthetic skill data: synthetic fine-tuning improves in-distribution retrieval but causes catastrophic forgetting on real and OOD skills. Continual-learning mitigations (embedding-anchor regularization, LwF, EWC, L2-init) keep OOD retrieval and add +13.98% on synthetic in-distribution for a 0.6B Qwen retriever+reranker. Named benchmark plus a recipe for scarce multi-positive supervision; EMNLP Industry 2026. Submitted Sep 9; abs verified. (arXiv cs.IR)
  • T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks (arXiv 2609.11042) — A 122B MoE trained to drive a real shell in a cloud sandbox for up to 300+ tool-call turns, rewarded by executing each task’s own verifier. The transferable parts are the stability tricks — warm-started actor-critic, dense pass-count process reward, TITO token-identity training with drift repair at turn boundaries, and rollout routing replay that records per-expert MoE choices and replays them at train time (train→inference logprob gap 0.021→0.013, zero token drift in the loss region) — plus a deliberately OOD corpus disjoint from Terminal-Bench 2.1 (base 43.8% → 64.0% resolved; Long-Horizon Terminal Bench 27.9%, ahead of GPT-5.4 and GLM-5.1 on the authors’ runs). Submitted Sep 10; abs verified. (arXiv cs.LG)

Models & research

  • SWE-2 — Cognition — Day-1 coding model post-trained from Kimi K3 (2.8T), with RL that trains all reasoning-effort levels in a single run under per-effort cost penalties: 50.0% on FrontierCode 1.1 Main (within 1 pt of Fable 5.1 while claiming 64% lower cost), DeepSWE 1.1 73.0% (above Fable 5.1’s 67.4%, just under Astra’s 74.1%), Terminal-Bench 2.1 92.8%, and Terminal-Bench 4 27.3% — a large drop that the post does not hide. Deployed in Devin Desktop/CLI/Web and Fusion; not open weights, and every number above is vendor-run. Median steps on FrontierCode drop 127 → 53 at medium effort, which is the efficiency claim worth testing against your own tasks. (HN 419 · 175 comments)
  • Structured Transforms for Low-Overhead Quantization of Language Models (arXiv 2609.11687) — Kashin-decomposition quantization revised: the dense random orthogonal transform becomes a sign-randomized DCT (per-iteration cost O(N²)→O(N log N)), the greedy alternating update gets closed-form cluster initialization (no k-means restarts), and combined with OPTQ-style error compensation it is competitive with OPTQ/QuIP at 4-bit on OPT/Llama-2/Pythia. The robustness result is the interesting one for self-hosters: where QuIP variants diverge to four-digit perplexity (Pythia-6.9B) or abort with NaNs (Mistral-7B), Kashin-DCT stays numerically stable near the FP16 baseline. Submitted Sep 10; abs verified. (arXiv cs.CL)

Industry

  • OpenAI pauses new $200/month ChatGPT Pro subscriptions — First-party post from an OpenAI staffer: new Pro signups paused amid “unprecedented” Astra demand, with existing accounts, other plans and the API unaffected. The capacity datapoint is the story — it lands the same week as the Astra rollout, and matches the compute-constraint reporting from the funding cycle. (Techmeme · X @thsottiaux · Sep 10)

Policy & provenance

  • OpenAI is open to slowing frontier development — and asked Congress whether an industry-wide slowdown is legal (Bloomberg) — Sources-say on both sides, paywalled, no artifact: Altman reportedly told staff OpenAI is weighing slowing its cutting-edge work and hopes rivals follow, and the company has asked lawmakers about the antitrust implications of coordinating an industry slowdown. Reported as reported — the substantive tension is that a slowdown is only meaningful if competitors can legally do it too.
  • Continued: Anthropic’s September 2026 threat-intelligence report — day 2 of coverage (base specs in yesterday’s digest) (full PDF · WSJ on the distillation section). What’s new: this is a separate 8-month retrospective (Dec 2025–Aug 2026) across seven harm areas — cyber operations, influence ops, surveillance, scams/fraud, biological misuse, conventional weapons, and illicit distillation, the last being the source of the WSJ story on Moonshot/DeepSeek routing user queries to Claude through “transfer stations” outside China. New safeguard details disclosed: reasoning-context editing blocked for new API accounts, identity verification for accounts in unsupported countries. No Fable/Mythos-class misuse except one distillation case. First-party account, vendor-framed; the PDF is the checkable artifact. (lobste.rs · HN 138 · 201 comments)
All gathered items - what was cut and why (16)