Google launches the score-only Gemini Decisions API, Microsoft finds editing under 10 neurons can backdoor an agent, AWS releases a local emulator
Today's thread: the small jobs agents do are getting cheaper and safer. Google replaces generative calls for routing and checks with a single forward pass, and AWS lets coding agents fail locally in an emulator and in x86 microVMs. But Microsoft's Lazarus is a reminder that weights are part of the supply chain too: where a model came from, and whether it was changed, needs checking like code.
One request takes a single multimodal input (text, images, audio, video and PDFs, up to 256K tokens) plus up to 1,000 independent questions, as yes/no, multiple choice (2 to 255 options) or ordered rubric scores, each answer with a confidence score from 0 to 1.
The model scores options in a single forward pass without token-by-token decoding, at 150 to 250 ms end-to-end; it costs $0.10 per million input tokens with no output tokens billed, and the shared input is billed once per request.
Google says it is nearly 5x cheaper across its shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API; Gemini Notebook uses it as a real-time router, serving everyday queries with up to 40% lower latency.
Integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI and others, with support coming soon to Gemini in Google Cloud.
Builder's takeRead with the 11-11 item on Hugging Face's common interface for decision models: picking from options is being split out of generative calls. I would move two kinds of calls first, grading interview answers against a rubric in AI Interview and pre-publish policy checks in PandaClaws, set a confidence threshold, and send only low-confidence cases to a large model or a person, then compare accuracy and cost per call before switching.
The researchers find that reprogramming fewer than 10 rarely firing neurons is enough to backdoor a model, with no training, needing only inference-level compute and write access to a checkpoint.
The trigger can be mundane, such as the clock passing midnight into 2027: mid-task, the agent then emits a command that downloads and runs the attacker's code, with whatever access it was given.
Frontier agents can spot anomalous weight edits when asked to look, but no scan can prove a model is safe; reliable detection requires diffing against the original weights, so provenance and integrity are what matter. The author stresses this is not a case against open weights: internal models are equally exposed.
Microsoft says Foundry ties each onboarded open-weight model to its original publisher and runs it in separately scanned containers, and it is building agent-based weight scanning, red-teaming tools and cryptographic model signing with verifiable provenance.
Builder's takeTwo things I would do right away: in the deploy pipeline, check self-hosted open-model weights against the publisher's hashes and give only the release account write access to the bucket that holds them; and default agent environments to an egress allowlist, so even a tampered model cannot complete the download-and-run step.
Existing AWS SDK and CLI calls go to the emulator and return in milliseconds; for supported operations, valid calls return the same result as AWS and invalid ones the same errors, while unsupported operations return 501 Not Implemented.
State lives in memory until you stop it; each copy gets its own port, keys and in-memory state, so agents, test suites and CI jobs never see each other's resources.
Services are generated from the API models AWS publishes, the same ones behind the AWS SDKs, so new services and APIs show up release by release, with a fidelity report comparing responses to AWS.
It does not depend on a Region and has no additional service fee; you pay only for resources when you deploy, and the source is published on GitHub under CC0 1.0 Universal.
Builder's takeI would point coding agents at the emulator for integration tests by default, one copy per task, and stop handing them real account credentials. Before relying on it, check the fidelity report for the services you use, and keep a smoke test against real AWS for anything that returns 501.
Built on Qwen3.5-0.8B, it improves on its base from 23.4 to 44.0 on BFCL v4, 30.1 to 58.7 on ACEBench and 23.2 to 51.5 on τ-bench.
It beats Qwen3.5-2B-Think on five of six tool-use and agentic benchmarks, trailing only on Claw-Eval (46.5 vs 48.9).
Its rate of getting stuck in repetitive reasoning loops on IFBench falls from 22.9% to 1.4%; lower-resource languages such as Hindi (24.5) and Russian (57.6) remain far weaker than English and Chinese.
The team says it will release open weights under Apache 2.0 for research and educational use, in bf16, Q8 and Q4, as safetensors and as GGUF files for llama.cpp.
Builder's takeNote the license is stated as research and educational use, so check before shipping it commercially. I would try it on single-step tool calls in the AI Cloud Drive mobile app, like finding files by condition or creating reminders, keeping those off the cloud; multi-turn tasks with complex rules still go to a cloud model, since that is where it drops off.
When creating a MicroVM image from a Dockerfile you can choose x86 or Arm (AWS Graviton, still the default), so existing x86 applications, tools and dependencies run without recompiling.
Lambda MicroVMs provides VM-level isolation, near-instant launch and resume, and state preservation, with listed uses including AI coding assistants, CI/CD runners, interactive development environments and vulnerability scanners.
x86 is available in all Regions where Lambda MicroVMs is offered except Asia Pacific (Hong Kong), Asia Pacific (Malaysia), Asia Pacific (Thailand) and Europe (Spain), through the console, CloudFormation, CDK or the Agent Toolkit for AWS.
Builder's takeTogether with Nitro Sandboxes on 11-10, AWS now offers isolation for agent code across containers, pods and functions. If x86-only binary dependencies kept you off MicroVMs, try it now; first estimate resume counts and cost for sessions that run briefly and wait long.
Content Understanding adds the GPT-5.6, GPT-6 and GPT-6.1 families; for standard document extraction Microsoft suggests starting with GPT-6.1 Sol, and GPT-6 Luna when token cost matters most.
In GA extraction, Astra leads Sol 95.95% to 95.74% with schema only, at about five times Sol's completion-token expense; in the Preview extraction evaluation Sol leads instead (96.06% vs 95.58%).
Adding five labeled examples lifts Luna's extraction quality from 91.82% to 94.16%, with token expense still much lower than Sol's.
The best model varies by task: Astra leads document segmentation (98.00% vs 97.10%), and GPT-6 Sol has the highest headline video Generation F1 (86.40%).
Builder's takeWhat this data shows is that the pricier tier often buys only a few tenths of a point. For field extraction in AI Cloud Drive, I would add five labeled examples to the cheap tier first and only then decide whether to upgrade, and evaluate and choose models separately for segmentation and extraction.
Multi-vector models keep one vector per token, so a corpus of millions of documents can hold hundreds of millions of vectors; token pooling merges similar vectors to shrink storage.
At pool_factor=4, a normally trained model retains 94.6% of its unpooled nDCG@10, and pooling-aware training raises that to 97.8%, with nDCG@10 rising from 49.03 to 51.23 while storing 74.8% fewer document vectors.
With adaptive pooling at an average of 29.58 vectors per document (84.7% fewer), the pooling-aware model retains 97.3% versus 93.3% for the normally trained one.
The method is implemented as a loss wrapper in Sentence Transformers, with a short reproducible training example.
Builder's takeIf AI Cloud Drive's semantic search moves to multi-vector models, storage cost is the main worry, and this post gives a reproducible trade-off. I would estimate vector counts and storage cost at pool_factor=4 first, then check recall on our own documents rather than relying on NanoBEIR.
Adapters convert the messages, tool calls and agent-to-agent handoffs of LangGraph, LangChain, Strands Agents and Bedrock AgentCore into one canonical trace, so every later step is framework-independent.
It has 36 built-in metrics across 7 dimensions: tool calling, response quality, responsible AI, performance, multi-turn conversation, multi-agent coordination and reasoning; tool-calling metrics need ground truth with the expected calls, order and arguments.
Traces are screened for PII such as emails, phone numbers and credit card numbers before storage, and runs are stored as experiments in DynamoDB and S3 for comparing versions.
AWS says 12 organizations in telecom, healthcare and financial services used it to compare agent versions before release, with runnable notebooks for each supported framework.
Builder's takeThe idea is worth copying even if the system is not: settle on one framework-neutral trace format and have evals read only that. I would first write a test set with expected tool-call order for the PandaClaws publishing agent and run it before and after any model or prompt change, which fits with the NeMo failures-to-tests loop from 11-05.
It is now a conditional model: a million-letter DNA sequence plus one target condition (tissue, cell type or state), picked from a reference catalog or generated on the fly from a scientist's own single-cell experiment, without retraining.
It is trained on bulk and single-cell data from human and mouse, including over 107 million individual cell profiles across 68 tissues, and adds DNA methylation as a readout.
On TenK10K, a single-cell map of 28 immune cell types, it identifies 30% of causal variants versus 16%, and 43% of directional gene effects versus 23%.
It is available today for non-commercial research via the AlphaGenome API, and in private preview for commercial research.
Builder's takeThis is far from my products; I include it for the design idea: make the target a conditional input instead of producing a fixed output per case. Products with many tenants or roles, such as AI Interview giving feedback per job, can try one model plus a condition before maintaining a prompt per role.
Researched and drafted with AI assistance; editorial standards and views set by Darius.
Every Monday, the week in AI in one email
Every Monday morning: the 6 AI stories that mattered most in the past seven days, with my take, in your inbox.
Your email is only used to send the weekly and is never shared. Every email has a one-click unsubscribe link. See the latest issue →