Engineers & builders · The Build
Every AI tool, API and release, the minute it ships.
Release radar
From official changelogs, GitHub releases and lab announcements.
OpenAI says the two models improve professional outputs, coding, computer use and efficiency, at lower cost.
The arXiv paper, submitted to ACM Computing Surveys, finds risk grows with agent authority and that most defences rely on model judgement and cover only isolated threats.
The expansion lets QPU builders write one file to connect hardware to any supercomputing center, and it supports the QRMI and QDMI interfaces.
Advisor now flags WorkSpaces problems without an admin starting an investigation, and it guides end users through fixes that IT has approved. It requires WorkSpaces client 5.37 for Windows.
The 5.24 billion parameter OpenTSLM-Flamingo runs on a single 24 GB GPU and takes numeric series plus text context, with Nexthink as a customer example.
- 13m agoAmazon SageMaker AI adds KV-aware routing, cutting p50 time to first token by up to 97 percent
AWS says the new RoutingStrategy sends each LLM request to the instance holding the longest cached prompt match, using cache events the vLLM container reports.
The action now detects interactive or automation mode on its own, and v0.x inputs such as mode and direct_prompt are replaced by prompt and claude_args.
- 14m agoAWS adds NVIDIA Dynamo to SageMaker HyperPod Inference, with 2.05x throughput in disaggregated test
The managed add-on supports multi-node serving, split prefill and decode, and KV-aware routing with vLLM, SGLang or TensorRT-LLM, with vLLM as the default engine.
The release also fixes the session tool runner and EnvironmentWorker so they stop running tools for terminated sessions.
Tool stack map
Agents that read your repo, edit files, run tests and open pull requests. Claude Code (66%) and GitHub Copilot (59%) were the most used coding agents in the 2026 Stack Overflow survey. Most teams run more than one. Commit an AGENTS.md or CLAUDE.md with build commands and conventions; GitHub found new repos with an agent configuration file logged 8x as many commits.
First-party APIs get new models and features first. Cloud marketplaces (Bedrock, Vertex AI, Foundry) trade a short lag for your existing billing, IAM and data-residency setup. Pin model IDs, read the deprecations page monthly, and keep a second provider wired up for failover.
Libraries for the loop of plan, call a tool, observe, repeat. LangChain (30%) and the OpenAI Agents SDK (24%) led the 2026 Stack Overflow survey. Start with your provider's SDK and a plain loop; adopt a framework when you need durable state, multi-agent handoffs or tracing you'd otherwise build yourself. Note that OpenAI's Agent Builder shuts down on 2026-11-30.
The Model Context Protocol (MCP) is the common way to expose tools and data to AI clients: write a server once and Claude, ChatGPT, IDE agents and others can call it. A2A covers agent-to-agent messaging. Treat every tool result as untrusted input: prompt injection through tool output is the main security risk.
Run open-weight models on a laptop or your own servers for privacy, offline use or predictable cost. Quantized models trade some quality for memory. Check the license before shipping: open weights often come with use restrictions.
If you can't measure it, every model upgrade is a gamble. Build a task-specific eval set from real traffic, grade it automatically, and trace production calls. OpenAI's Evals dashboard and API go read-only on 2026-11-30, and OpenAI's docs point users to Promptfoo.
Retrieval quality decides RAG quality. Hybrid search (keywords plus embeddings) with a reranker beats either alone on most business documents. If you already run Postgres, start with pgvector and move only when you measure a reason to.
Fine-tune for format, tone and narrow tasks after prompting and retrieval stop improving your evals. Most teams fine-tune open-weight models with LoRA. Plan around OpenAI's timeline: new organizations lost self-serve fine-tuning on 2026-05-07, and existing customers can no longer create new jobs from 2027-01-06.
For self-hosting at scale, the open-source servers handle batching, KV-cache management and speculative decoding for you. Managed inference providers serve open-weight models per token. Compare on your own prompts: throughput and latency numbers vary a lot with input length.
For workflows that glue SaaS apps together with an AI step in the middle. Make (39%) and n8n (33%) led automation tools in the 2026 Stack Overflow survey. Put a human approval step before anything that emails customers or moves money.
Open source
Contributors recorded their own unscripted work with a head-mounted phone camera, and 83% of the hours come from workplace settings, under a CC BY 4.0 license.
The 5.24 billion parameter OpenTSLM-Flamingo runs on a single 24 GB GPU and takes numeric series plus text context, with Nexthink as a customer example.
- 14m agoStudy sets exact bounds on how model edits weaken OpenStamp watermarks in Llama-2-7B and Mistral-7B
INT8 quantization and 30% pruning keep detection at 85.6-94.8%, but 500-step LoRA fine-tuning cuts it to 28.8% and 11.6%, the authors report.