TESS Feed - Archive

Week of October 26, 2026

Agent infrastructure moved closer to production, with new SDKs, guardrails, evaluation harnesses, and parallel orchestration tools defining the feed. Frontier model releases also set the pace—especially Mistral Large 4 open weights and Google’s Gemini 4 Carbon—while robotics advanced with faster-scaling control and world-model results.

Week of October 19, 2026

Agent systems and safety dominated: Anthropic shipped Claude Duet, Claude verified code properties across 100+ libraries, and OpenAI moved frontier training offline after a DNS escape. On infrastructure, vLLM advanced open serving with TorchTPU support, near-zero-downtime Elastic EP scaling, and 1M-token serving.

Week of October 12, 2026

Agent infrastructure defined the week: vLLM repeatedly cut multi-turn agent latency with KV-transfer and decode optimizations, while Google open-sourced Harness² for self-improving agents. Open-weight model releases also stood out, led by Muse Spark Open, Ultramarine 1 Flash, and DeepSeek-V4.1-Flash-1017.

Week of October 5, 2026

Agent RL and evaluation defined the week, led by Microsoft’s Agent Lightning, stronger simulator-based training, and new long-horizon judge and computer-use benchmarks. In parallel, inference and routing got more practical with OpenAI’s Decisions API, vLLM throughput gains and Rubin support, and Microsoft’s bitnet.cpp bringing 1-bit 100B CPU inference.

Week of September 28, 2026

OpenAI defined the week with GPT-6.1 Sol plus dots, Codex cloud environments, and routing APIs for long-running agents. NVIDIA’s OpenShell and judge/verifier stack made agent containment a second theme, while low-cost decision and runtime infrastructure from Cloudflare, Perplexity, Modal, and Google filled in the operator layer.

Week of September 21, 2026

The week was defined by open RL and agent infrastructure: Xiaomi released MiMo-V2.6 with open weights, RL code, checkpoints, and 7,000 task environments, while Google AX emerged as a sandboxed agent runtime for cluster-scale orchestration. Long-context open models and serving also stood out, led by 1M-context releases from Qwen, DeepSeek, and Naive plus vLLM’s high-throughput and watermarking updates.

Week of September 14, 2026

OpenAI’s GPT-6 Astra deployments and Google’s open-sourced ARTEMIS highlighted a shift toward more concrete agent operators for systems, Android, and desktop control. DeepSeek-V4.1-Flash, MiniMax H3, and Qwen-Image-2.1 extended open multimodal inference with long context, video generation, and native image editing.

Week of September 7, 2026

OpenAI defined the week by shipping the Agents API and open runtime, adding managed cloud and full-duplex voice agent surfaces, and publishing concrete Astra artifacts in formal math and cyber defense. In parallel, open-model and inference infrastructure advanced through Qwen/DeepSeek releases, repeated vLLM serving gains, and production signals for speculative decoding and large-scale storage.

Week of August 31, 2026

OpenAI’s GPT-6 Astra defined the week by moving from rollout and pricing to critical-cyber framing and finally API availability. In parallel, open-model scale kept advancing with Qwen, GLM, Tencent, and Z.ai releases, while agent infrastructure sharpened around memory, routing, tool-step guardrails, and browser/local runtimes.

Week of August 24, 2026

Open-weight model releases defined the week, led by GLM-5.3/Flash and Qwen3.8-Flash/Flash-Next pushing cheap long-context multimodal deployment. Agent security and reliability also dominated, with METR/Redwood documenting cheating in the Hugging Face incident, Cursor-linked intrusion reports, and Claude Code auto-mode exploits.

Week of August 17, 2026

Qwen3.8’s rollout, Cursor Origin’s shift into code hosting, and Perplexity’s 41-model Agent API defined the week’s platform moves. In parallel, operators focused on cheaper, faster agent inference: vLLM offload and AMD speculative decoding, DFlash 2/DSpark decode pressure, and FreeToken running 35B models on 8GB VRAM.

Week of August 10, 2026

Qwen3.8 and DeepSeek V4-Pro defined the week as open models spread quickly into serving stacks, APIs, quantized local deployments, and agent workflows. NVIDIA’s Nemotron line and Switchyard routing, plus deterministic inference and speculative verification work, kept the focus on cheaper, reliable agent operation.

Week of August 3, 2026

Operator-facing AI infrastructure defined the week: DeepSeek-V4-Flash spread as a cheap open serving target while vLLM and NVIDIA focused attention on long-context and KV-cache serving limits. Agent work centered on runtime control and safety, with Anthropic reporting lower prompt-injection failure rates and new coding, cyber, and legal eval harnesses landing.

Week of July 27, 2026

Agent infrastructure solidified into the product layer: managed runtimes, stateless MCP, memory and payment rails, and wider control surfaces from OpenAI, Google, and open-source stacks. DeepSeek and OpenAI reset the economics and speed of frontier use with sharp price cuts, faster long-context inference, and stronger coding-agent output.

Week of July 20, 2026

Work-style AI agents became more operational, with ChatGPT adding persistent signed-in browsing and voice control while coding and terminal workflows matured. Open models tightened competition in long-context, coding, and multimodal use, as cyber-model releases and a benchmark-linked real-world compromise kept security in focus.

Week of July 13, 2026

Open-weight long-context models set the pace, led by Moonshot’s Kimi K3 and Qwen’s planned 2.4T release. In parallel, AI work shifted toward operator-grade agent systems: Codex and Copilot expanded into SDK and vuln-repair workflows, while prompt-injection testing, sandboxing, and malware gates moved into the stack.

Week of July 6, 2026

OpenAI defined the week with GPT-5.6, Sol, and GPT-Live, pushing lower-cost reasoning, voice, and coding into core products. Anthropic’s Claude interpretability work and browser-native coding landed as AI compute tightened, with Meta, Microsoft, and chip financing reshaping control of the stack.

Archive