Run agents more robustly, automate decisions: new AWS features, Gemini Decision, evaluation

Important AI news

AWS: cheaper models, steadier agents, JSON without cleanup

AWS bundles the October updates: cheaper models, more mature agent tools, and structured outputs. According to Anthropic, Claude Haiku 5.5 costs around 75 percent less than Haiku 4.5 for many tasks. OpenAI GPT-6.1 Terra is now generally available on Bedrock. According to OpenAI benchmarks it is close to GPT-6 Astra, runs in parts over three times faster than GPT-6 Astra and GPT-6.1 Sol, and has at least 50 percent lower token prices than GPT-6.1 Sol and GPT-5.6 Terra. Also new: Nova 2.5 Sonic for speech, GLM 5.3 for coding, and TwelveLabs Pegasus 1.5 for video. More models arrive in GovCloud. Important for operations: AgentCore CLI v1.0, Gateway with Bedrock Invoke APIs and model fallback, Step Functions as tools, and structured outputs for all Claude models on Bedrock. Numbers on price and quality come from the vendors. For you this means: you can define fallbacks, plug existing workflows in as tools, and process JSON directly.¹

Google launches Gemini Decision 1

Google ships Gemini Decision 1, a decision model that processes text, images, audio, video, and PDFs. It supports yes/no, selection, and rubric scoring. Price per source: 0.10 US dollars per one million input tokens, output free. Jev stays cheaper (0.042 US dollars) but is text based. In tests, Gemini returned probabilities instead of hard 0/1 values. You can standardize review decisions with it, such as moderation, visual checks, or ticket routing across media formats.²

23 reusable GenAI assets from AWS

AWS offers a curated catalog with 23 production ready building blocks, such as RAG modules, agent templates, and evaluations. Case studies report notable effects: DoorDash’s voice contact center with 2.5 second latency and 50 times test capacity, Ricoh’s document solution with seven times throughput and over 90 percent fewer engineering hours per deployment. These numbers come from AWS case studies. Benefit for you: you start with vetted blocks instead of zero and save hardening and security reviews.³

Unified evaluation for agents: UAEF

The Universal Agent Evaluation Framework (UAEF) evaluates agents in a framework agnostic way. Adapters normalize traces from LangGraph, LangChain, Strands Agents, and Bedrock AgentCore into a common format. 36 integrated metrics cover seven quality dimensions, results land in DynamoDB and S3. Twelve organizations already used it for version tests before releases. If you build agents, you get consistent metrics and spot regressions early.⁴

Compare policies semantically instead of text diffs

An open approach shows how to compare formal policies semantically. Core is anti unification. That means finding the most specific common pattern of two expressions. An e graph records known equalities. You see shared parts once and only true differences twice, for example in Cedar, Dogwood, OpenShell, or CEL. For autoformalization this helps with diverging LLM translations. Optionally, SMT solvers provide counterexamples. You reduce review risk when policies change.⁵

Highlights for your workday

• Keep agents highly available. Define a fallback chain across multiple models or providers in the AgentCore Gateway. Plug existing Step Functions workflows in as tools so that processes like refunds run under control. Structured outputs from Claude save you post processing.¹
• Scale decisions. Use Gemini Decision 1 for yes/no checks, selections, and rubric scores across text, images, or video. Pick Jev if text is enough and cost comes first. Plan the answer format as JSON from the start.²
• Do not reinvent it. Search the 23 AWS assets for contact centers, document processing, or evaluation. Start with a demo deploy per guide and measure a target metric, such as latency or field accuracy. Note: performance figures come from vendor case studies.³

Tools and updates

• AWS Emulator: official local emulator from AWS for tests without paid keys. Fits CI and developer tests against AWS APIs. Requirement: Docker on macOS or Windows.⁶
• n8n 3.2.0: raises the default assistant to Claude Opus 5.5 and adds Microsoft Work IQ OAuth2. Relevant for teams with automation workflows. Update to 3.2.0 required.⁷
• Aq TUI for Claude Code: terminal UI to control multiple Claude Code sessions in parallel, including review loops and an “Attention Queue”. Requires tmux and hooks to Claude Code.⁸
• BotPsy on device assistant: personal assistant with a small speech model on device. Useful for private notes and tasks without internet. Check device support first.⁹

Try it in five minutes

1) Define decisions. Take three sample support emails and write a short rubric with three criteria from 0 to 2 points. Set two more outputs: escalate (yes/no) and team (selection).² 2) Ask your existing AI chat tool to score the three emails by the rubric and return one JSON per email. Schema: {"rubric": {k1, k2, k3}, "score": 0–6, "escalate": true/false, "team": "Billing|Tech|Ops"}. Tip: demand strictly valid JSON. Structured outputs are also available via API on Bedrock with Claude.¹ 3) Check value. Does the model count consistently, does the escalation logic match the score threshold, and is the JSON readable without cleanup? If yes, you can map the fields into your ticketing system.²

A good find

Beep Locator: a pragmatic project that finds the source of an annoying beep by triangulation with multiple devices. Takeaways: you can measure with everyday devices, detect events with a simple FFT detector, and align recordings with a reference chirp. A clap test function built trust. Use the idea for alarms in warehouse, server room, or office.¹⁰

Sources

1. ICYMI: What landed for AI builders in October 2026 (aws.amazon.com)
2. Gemini Decisions API (simonwillison.net)
3. Accelerating generative AI delivery with reusable assets (aws.amazon.com)
4. UAEF: A framework-agnostic system for closed-loop evaluation of AI agents (aws.amazon.com)
5. Semantic diffing of formal policies with anti-unification (aws.amazon.com)
6. aws-emulator (simonwillison.net)
7. n8n@3.2.0 (github.com)
8. Show HN: Aq – an attention queue TUI for Claude Code (news.ycombinator.com)
9. BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model (huggingface.co)
10. Locating the source of a mysterious beep with Claude (simonwillison.net)