{"date":"2026-11-17","title":"What happened in AI — Nov 17, 2026","generated_at":"2026-11-17T22:16:37+00:00","intro":["Coding-agent cost became measurable today. A 27,280-run benchmark found 64% of inference spend went to failed runs and that the best harness depends on the model, while Microsoft's TokenEconomizer trims what agents send per turn.","Around that, the production layer filled in: Datadog now tests agent versions on live traffic, LangSmith restricts trace access, and shared memory and messaging tools (Loom, Woolly, Pulsar) span harnesses. OpenAI shipped GPT-6.2 Astra and Sol."],"highlights":["64% of spend in 27,280 coding-agent runs went to failures; re-pick your harness when you switch models.","Datadog Online Experiments compares agent versions on real production traffic, not just eval sets.","LangSmith adds access controls for sensitive data inside agent traces.","ClickHouse's open-source Loom gives coding agents shared code memory (0.78 → 0.84 answer quality).","SageMaker endpoints now route requests to the instance with the longest KV-cache prefix match.","OpenAI released GPT-6.2 Astra and Sol across ChatGPT, Codex, and the API."],"article_count":18,"categories":[{"name":"Coding-agent spend: harness choice decides the bill","slug":"coding-agent-spend","summary":"Two large measurements say the harness and the context it sends, not just the model, decide what a coding agent costs and whether the money buys a passing run.","articles":[{"title":"In 27,280 coding agent runs, 64% of inference spend went to runs that failed","summary":"FrontierHarness v1.1 ran 22 harnesses × 5 models × 62 tasks × 4 attempts (83B tokens, with Baseten): 64% of inference spend went to failed runs, and the best harness changes with the model.","source":"hackernews_ai","url":"https://runta.com/blog/introducing-frontierharness-eval-v1-1/","published":"Tue, 17 Nov 2026 14:53:43 +0000"},{"title":"TokenEconomizer: Learning from Experience to Make AI Coding Agents Efficient","summary":"Microsoft Research's TokenEconomizer learns from past sessions which instructions, tool descriptions, and history an agent actually needs, cutting tokens sent per coding-agent turn.","source":"hackernews_ai","url":"https://www.microsoft.com/en-us/research/blog/tokeneconomizer-learning-from-experience-to-make-ai-coding-agents-more-efficient/","published":"Tue, 17 Nov 2026 06:53:51 +0000"},{"title":"NOX Local – a coding agent for 4B models on an 8 GB GPU","summary":"NOX Local builds the whole agent loop around a 4B model on llama.cpp and an 8 GB GPU, with MCP, skills, and plan mode, and no API key or per-token bill.","source":"hackernews_ai","url":"https://github.com/Johnny008-creator/nox-local","published":"Tue, 17 Nov 2026 12:46:55 +0000"}]},{"name":"Agent memory and multi-agent coordination","slug":"agent-memory-coordination","summary":"Memory and messaging are moving out of individual harnesses into shared, open layers that every coding agent on a machine or team can use.","articles":[{"title":"Loom: agent memory that keeps up with your code","summary":"ClickHouse open-sourced Loom, a shared code index plus learned-notes store for coding agents; it raised mean answer quality on ClickHouse code questions from 0.78 to 0.84.","source":"hackernews_ai","url":"https://clickhouse.com/blog/introducing-loom-agent-memory-that-keeps-up-with-your-code","published":"Tue, 17 Nov 2026 10:20:15 +0000"},{"title":"Show HN: Woolly, a local mem0-compatible memory API for AI agents","summary":"Woolly is a local, mem0-compatible memory API that keeps one readable Markdown memory across Cursor, Claude Code, and Codex via SDKs, plugins, and MCP.","source":"hackernews_ai","url":"https://usewoolly.com/","published":"Tue, 17 Nov 2026 04:01:19 +0000"},{"title":"Show HN: Pulsar – cross-harness agent to agent communication and messaging","summary":"Pulsar is an MCP server that puts every local coding agent (Claude Code, Codex, Cursor, Gemini CLI and others) on one bus for direct, tagged, or broadcast messages.","source":"hackernews_ai","url":"https://github.com/yocontra/pulsar","published":"Tue, 17 Nov 2026 03:18:53 +0000"},{"title":"I Built My Own AI Agent Orchestration Tool","summary":"Checkly's Agentic Crafting Table maps 10–20 concurrent coding agents so the operator sees which one needs input instead of polling each session.","source":"hackernews_ai","url":"https://www.checklyhq.com/blog/agentic-crafting-table-ai-agent-orchestration/","published":"Tue, 17 Nov 2026 13:00:52 +0000"},{"title":"What We’ve Learned Building Proactive Agents","summary":"LangChain's lessons on proactive agents: grant autonomy in stages like a new hire, from scoped tasks to open goals with fewer check-ins.","source":"langchain_blog","url":"https://www.langchain.com/blog/what-weve-learned-building-proactive-agents","published":"2026-11-17T17:00:00+00:00"}]},{"name":"Running agents in production: observability, access, runtimes","slug":"agents-in-production","summary":"Agent platforms are adding the controls platform teams need to run agents at scale: test on live traffic, restrict who sees trace data, and pick the right runtime lifecycle.","articles":[{"title":"Datadog Brings Production Experimentation to AI Agent Observability","summary":"Datadog Agent Observability adds Online Experiments, comparing agent versions against real production traffic instead of only fixed eval datasets.","source":"infoq_ai_ml","url":"https://www.infoq.com/news/2026/11/datadog-ai-agent-observability/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering","published":"Tue, 17 Nov 2026 12:00:00 GMT"},{"title":"Protect sensitive data in your traces: new controls in LangSmith","summary":"LangSmith adds three controls that limit who can see sensitive data inside production agent traces, so engineers can debug under least-privilege rules.","source":"langchain_blog","url":"https://www.langchain.com/blog/protect-sensitive-data-in-your-traces-new-controls-in-langsmith","published":"2026-11-17T18:00:00+00:00"},{"title":"Why Run AI Coding Agents on Kubernetes?","summary":"Okteto argues coding agents should move from laptops to Kubernetes once platform teams roll them out to hundreds of developers, for isolation and governance.","source":"hackernews_ai","url":"https://www.okteto.com/blog/why-run-ai-coding-agents-on-kubernetes/","published":"Tue, 17 Nov 2026 16:37:26 +0000"},{"title":"Not Every Agent Hibernates","summary":"Of four common agent patterns, only two ever suspend, so hibernation is not a fleet-wide fix for idle-agent cost; the rest are torn down or run to completion.","source":"hackernews_ai","url":"https://frankdenneman.ai/2026-11-17-not-every-agent-hibernates/","published":"Tue, 17 Nov 2026 11:39:00 +0000"}]},{"name":"Inference serving: cache routing and latency budgets","slug":"inference-serving","summary":"Serving work today targets prompt-cache reuse and tight tail latency, plus a new open runtime for non-NVIDIA silicon.","articles":[{"title":"Amazon SageMaker AI real-time endpoints now support KV cache aware routing","summary":"SageMaker AI real-time endpoints can now route each request to the instance holding the longest cached prompt match, with guidance on when to prefer prefix-aware routing.","source":"aws_ml_blog","url":"https://aws.amazon.com/blogs/machine-learning/amazon-sagemaker-ai-real-time-endpoints-now-support-kv-cache-aware-routing/","published":"Tue, 17 Nov 2026 21:11:51 +0000"},{"title":"Presentation: Scaling LLM-Based Ranking Systems for Latency-Critical Search & Recommendation Workloads","summary":"LinkedIn meets P99 SLAs for LLM ranking by distilling 7B to 400M parameters and compressing job context into single-token embeddings for a 10x throughput gain.","source":"infoq_ai_ml","url":"https://www.infoq.com/presentations/scale-llm-ranking-systems-linkedin/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering","published":"Tue, 17 Nov 2026 11:00:00 GMT"},{"title":"Show HN: TrnRT – An open-source LLM inference engine for AWS Trainium","summary":"RunAnywhere open-sourced TrnRT, an LLM inference engine for AWS Trainium that already serves GLM 5.3 flash in FP8.","source":"hackernews_ai","url":"https://github.com/RunanywhereAI/trnrt","published":"Tue, 17 Nov 2026 16:01:04 +0000"}]},{"name":"Models: OpenAI ships GPT-6.2 Astra and Sol","slug":"models","summary":"OpenAI's 6.2 generation lands across ChatGPT, Codex, and the API, with a vertical legal variant on top.","articles":[{"title":"Introducing GPT-6.2 Astra and Sol","summary":"OpenAI released GPT-6.2 Astra and Sol, claiming better coding, computer use, and professional output at lower cost in ChatGPT Work, Codex, and the API.","source":"openai_blog","url":"https://openai.com/index/introducing-gpt-6-2-astra-sol","published":"Tue, 17 Nov 2026 17:45:00 GMT"},{"title":"Astra for Law leads in legal intelligence","summary":"Astra for Law ranked first in OpenAI's launch evaluation across 24 areas of law; ChatGPT Enterprise customers can request access.","source":"openai_blog","url":"https://openai.com/index/astra-law-6-2","published":"Tue, 17 Nov 2026 00:00:00 GMT"},{"title":"CLI for torrenting open weight AI models","summary":"Pirate Face CLI fetches 37,160 open-weight models over BitTorrent and falls back to Hugging Face, so a stalled 15 GB CI download no longer fails the job.","source":"hackernews_ai","url":"https://pirateface.co/articles/pirate-face-cli","published":"Tue, 17 Nov 2026 16:53:07 +0000"}]}],"video":{"youtube_id":"ZLu-4_vbJwY","title":"LLM Digest Daily Explainer: Nov 17, 2026","duration_s":116.7}}