Agents Take on Sustained Work as Human Control Faces New Tests

Thursday, 12 November 2026 | Global AI News Daily Ledger

Today's strongest signal is agents taking on sustained work. Amazon released a local AWS emulator it says was written and maintained entirely by autonomous agents, Databricks turned Genie into an always-on enterprise coworker, and Alibaba opened both its Qoder agent harness and a two-week preview of its Qwen4 flagship. Google shipped a purpose-built Decisions API for agent routing and grading, and a much sharper genomics model, while Runway put an open 27B robot model under Apache 2.0. The pushback is sharpening too: United moved to keep unauthorized AI agents from completing flight bookings, Helsing put hundreds of agents into military target preparation under a named-human rule, and Anthropic opened a private beta for hosting its agent service inside a customer's own environment while keeping model inference separate.

Coverage window Since 11 November, 18:45 IST | Compiled 12 November, 18:50 IST

1. Amazon releases AWS Emulator, which it says was written entirely by AI agents

AWS on 11 November released AWS Emulator, a free local emulator of its cloud services that runs on a developer's machine as a Rust binary, CLI or Docker container on Linux, Windows and macOS. Senior Principal Engineer Matt Goldberg says thousands of autonomous AI agents wrote every line of its code and continue to maintain it, with no human writing or reviewing any of it. Three agent teams divide the work: one writes fidelity tests from public Smithy API models and verifies each against the live cloud, one replays those tests against the emulator and files grouped issues, and one fixes one issue per agent through an automated merge gate with security and fidelity checks. Goldberg says a single Kiro prompt produced a passable Amazon SQS replica in 15 minutes. The code is on GitHub under a CC0 public-domain dedication with a public "things-that-do-not-work" backlog, and the emulator redirects ordinary AWS CLI or SDK calls to localhost for free offline testing.

Analysis. AWS presents this as an infrastructure-software project written and maintained end-to-end by agents, and the load-bearing piece is the guardrail system - verified fidelity tests plus the merge gate - rather than the model. The caveat is fidelity: emulation of outages, limits and service-specific behavior remains incomplete, so it validates code paths, not production behavior.

https://www.aboutamazon.com/news/aws/how-autonomous-ai-agents-built-aws-emulator
https://siliconangle.com/2026/11/11/amazons-open-source-aws-emulator-lets-developers-validate-their-code-locally-without-the-cloud/
https://www.allthingsdistributed.com/2026/11/aws-emulator-and-the-engineers-job-when-agents-write-the-code.html

2. Google DeepMind's AlphaGenome 2 predicts what DNA variants do in individual cell types

Google DeepMind on 11 November announced AlphaGenome 2, a conditional DNA-sequence model that predicts how genetic variants affect a specific human cell type or state, rather than averaging across tissues. It was trained on more than 107 million individual cell profiles across 68 human and mouse tissues and adds DNA methylation to its predicted readouts. On the TenK10K dataset of 28 immune cell types, DeepMind reports it identified about 30% of causal variants versus 16% for a fine-tuned AlphaGenome 1, and predicted the direction of gene effects in 43% of cases versus 23%. The model computationally reproduced a known inflammatory bowel disease mechanism, isolating the distant regulatory variant that overactivates ETS2 in gut macrophages. It is available now for noncommercial research through the AlphaGenome API, in closed preview for commercial use, and is not approved for clinical diagnosis.

Analysis. Moving from bulk-tissue averages to single-cell resolution is a genuine architectural step for genomics AI, and the ability to condition on a few dozen cells from an unseen experiment without retraining is the part working researchers will test first. The performance figures are DeepMind's own; independent replication and any path to clinical use are still ahead.

https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/
https://the-decoder.com/google-deepmind-shifts-genomics-ai-from-bulk-tissue-predictions-to-single-cell-type-resolution-with-alphagenome-2/

3. Alibaba opens a two-week Qwen4-Max flagship preview, without publishing benchmarks

Alibaba on 11 November opened Qwen4-Max-Preview, the first public release in its Qwen4 flagship line, through its QwenCloud Token Plan only. Official documentation fixes the window from 10:00 on 11 November to 24:00 on 25 November (UTC+8), says the model will not change during it, and says it will be retired afterward; pay-as-you-go access is not supported. Listed specifications include a 1M-token context, a maximum output of 393,216 tokens shared between thinking and the final answer, thinking mode on by default, image and video understanding, function calling and built-in tools, and compatibility with OpenAI Chat Completions, Responses and Anthropic Messages APIs. A limited-time promotion prices credit consumption as low as 20% of the standard rate, dropping to 4% overnight. Notably, official materials for the preview include no benchmark chart or evaluation table.

Analysis. A frontier flagship preview with full specs but zero published evals is an unusual launch posture, and the hard two-week window could be a validation period, but does not establish a date or commitment for general availability. Token Plan-only gating keeps independent third-party testing thin until then.

https://www.alibabacloud.com/help/en/model-studio/qwen4-max-preview
https://digintoai.com/en/articles/qwen4-max-preview

4. Runway open-sources Praxis-1, a 27B robot control model, under Apache 2.0

Runway on 11-12 November released open weights for Praxis-1, a 27-billion-parameter mixture-of-experts diffusion transformer for robot control, under Apache 2.0 with commercial use permitted. The model pairs a 26.9B video backbone with a 16-layer action expert, activates roughly 2.8B parameters per token, and predicts camera video plus 24 action rows covering 1.6 seconds at 15 Hz per call; base, DROID and bimanual YAM variants and 4-bit NVFP4 builds are on Hugging Face. Runway reports 46.2% success across 1,200 episodes on NVIDIA's RoboLab-120 benchmark, ahead of FLUX 3 Action (42.9%) and Physical Intelligence's pi0.5 (28.0%), and says fine-tuning transferred to Google's RT-1 robot - an embodiment not in training - at 80.0% on SimplerEnv versus 52.4% for RT-1's original policy. The training mix is 58% egocentric human video, 20% general video and 18% real robot video.

Analysis. A commercially usable robot foundation model from a video lab, free under Apache 2.0, pressures every closed robotics stack to justify its premium. The scores are Runway's own, measured against other vendors' published figures, so the decisive test is third-party RoboLab-120 replication, not another vendor post.

https://runway.com/news/company-news/praxis-1-is-now-available
https://runway.com/research/praxis-1-video-pretraining-to-robot-control
https://huggingface.co/collections/runwayml/praxis-1
https://aiweekly.co/alerts/runway-open-sources-praxis-1-a-27b-robot-model-apache-20

5. Alibaba's Qwen Code 1.0 Preview open-sources the agent harness behind its Qoder tool

Alibaba on 11 November published Qwen Code v1.0.0-preview on GitHub, folding the agent harness it built for its commercial Qoder coding tool into the open-source Qwen Code terminal agent under Apache 2.0. The preview ships native macOS, Linux and Windows packages with no Node.js dependency, supports Qwen and Qoder accounts plus custom OpenAI-compatible endpoints, and adds session resumption and non-interactive runs for scripts and automation. GitHub marks it as a pre-release, the release notes warn that interfaces and behavior may change, and the team is temporarily not accepting community pull requests until after December 2026. The repository showed about 29,500 stars at release. Qwen Code originated as a fork of Google's Gemini CLI and has been developed independently since v0.1.

Analysis. Chinese labs are now productizing the agent layer, not just the model: DeepSeek has DeepSeek Harness, and Alibaba is giving away the runtime it sells inside Qoder so others embed it. An open harness with central MCP, permission and sandbox management is the piece enterprises actually build on.

https://github.com/QwenLM/qwen-code/releases
https://www.llms.blog/posts/qwen-code-1-0-preview-alibaba-opens-its-qoder-agent-harness

6. Google launches the Gemini Decisions API for fast, cheap multimodal choices

Google on 11 November introduced the Gemini Decisions API, powered by a new Gemini Decision 1 model, in Google AI Studio. Instead of generating text token by token, the endpoint scores up to 1,000 yes/no, multiple-choice and rubric questions against one shared multimodal input - text, images, audio, video and PDFs up to 256K tokens - in a single forward pass, returning probabilities and confidence scores. Google cites 150-250 ms end-to-end latency and pricing of $0.10 per 1M input tokens with zero output tokens billed, and says it is nearly 5x cheaper than Gemini 3.5 Flash-Lite across its shared benchmark suite. Google teams already use it for model routing in Gemini Notebook, rapid wireframing in Stitch, emoji selection in CC and adversarial screening of MCP tools in Search. SDK and ecosystem integrations span OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, with Gemini in Google Cloud support coming.

Analysis. Routing, gating and grading are the unglamorous control plane of every agent system, and Google is selling that as a purpose-built primitive instead of a prompt pattern. The latency, calibration and cost figures are vendor numbers, but the zero-output-token pricing model alone could attract routing workloads.

https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/
https://simonwillison.net/2026/Nov/11/gemini-decisions-api/

7. Databricks turns Genie into an always-on agent for long-running enterprise work

Databricks on 11 November launched the new Genie, repositioning its assistant as an autonomous AI coworker that runs continuously on delegated jobs - monitoring pipelines, chasing missing inputs and acting within set guardrails - reachable through chat, email, Slack and Microsoft Teams. The release bundles Genie Ontology, a business-context layer built mostly by agents and governed through Unity Catalog, plus Genie Search, a hybrid index-and-MCP search that enforces each source system's permissions, and Unity Gateway for admin budgets, tracing and tool access. MCP connectors let Genie act in Slack, Zoom, Microsoft 365, Jira and GitHub. In a vendor benchmark built with Turing, Databricks says Genie answered 91% of task runs correctly, 1.6x the rate of a stronger frontier agent, using less than half the tokens. CEO Ali Ghodsi showed TechCrunch his own production use, including employee-survey analysis and sales-call postmortems.

Analysis. The enterprise agent race is shifting from answer engines to delegated ongoing responsibility, and Databricks' bet is that governed context, not model quality, is the differentiator. The 91% figure comes from a benchmark Databricks built itself, and the dogfood demo is a CEO anecdote, so treat both as directional.

https://www.databricks.com/blog/new-genie-autonomous-ai-complex-enterprise-work
https://techcrunch.com/2026/11/11/databricks-new-ai-agent-tells-ceo-ali-ghodsi-the-unvarnished-truth/

8. United Airlines moves to block unauthorized AI agents from booking flights

United Airlines said on 11 November it will restrict AI bots' access to its digital platforms, citing data security. United's current policy also prohibits unauthorized agents from searching, managing or collecting information on its own site or app; general AI trip research is allowed only without accessing those platforms, and customers must complete purchases themselves, and the airline warned that sharing account credentials with third-party assistants risks exposing personal information, unintended bookings and unauthorized use of frequent-flyer miles. It also flagged regulatory exposure: check-in requires a human acknowledgment of prohibited items such as explosives. United said it will enforce the restrictions directly and indirectly through travel agencies and aggregators, updating its booking, ticketing and distribution policies. CEO Scott Kirby framed the move as refusing another middleman between the airline and its customers. The decision lands amid divergent industry responses: Amazon has blocked Meta's Muse agent while Walmart partnered with it, and Expedia and Booking are working with OpenAI.

Analysis. United's decision offers a template for agent-commerce pushback: incumbents control the checkout and the regulatory declarations, so assistants get to advise but not transact. Expect 'authorized agent' certification programs, not open access, to be the compromise.

https://www.cnbc.com/2026/11/11/united-airlines-ai-agents.html
https://www.united.com/en/us/fly/help-center/digital-trust-center/artificial-intelligence/third-party-ai-agents.html
https://www.washingtonpost.com/travel/2026/11/12/united-airlines-blocks-ai-agents-booking-flights/

9. Helsing's Clarity puts hundreds of AI agents into military target preparation

German defense company Helsing unveiled Clarity in Berlin, an 'agentic intelligence platform' in which hundreds of specialized AI agents fuse sensor feeds, open-source intelligence, written reports and existing systems to identify adversary forces, facilities and positions, then package targets for the appropriate weapons system. Helsing says a human operator must authorize every handoff to a weapon, targets are recommended only when independent sources agree with high confidence, and every decision is documented under a responsible person's name. The platform runs on customer infrastructure, implements NATO standards and is pitched as a sovereign European alternative to Palantir's Maven; Helsing did not say whether any customer has deployed it. Independent reporting notes the failure modes of comparable systems: reported 20-second reviews in Gaza's Lavender pipeline, 'the computer said to do this' attitudes around Maven in Iran, and a spring 2026 near-miss in which a chatbot's false claim about a Chinese ship's cargo nearly triggered a boarding.

Analysis. Agent swarms are moving from copilots to the targeting chain faster than oversight practice can follow. The named-human rule is the right design on paper, but the cited reporting on comparable systems shows how review time can collapse under operational tempo, which is exactly what outside scrutiny should watch.

https://helsing.ai/newsroom/helsing-unveils-clarity-a-sovereign-software-platform-for-military-intelligence
https://the-decoder.com/helsing-built-an-ai-system-that-prepares-targets-for-weapons-systems-and-says-every-decision-will-carry-a-human-name/

10. Anthropic opens self-hosted Claude Managed Agents in private beta

Anthropic on 12 November made self-hosted Claude Managed Agents available in private beta, with access requested through an account team. Customers can run the agent service, sandboxes, credentials and session history in their own cloud or on-premise, operating the server, database and workers themselves. Previously, customers could host the sandbox while Anthropic ran the rest of the service. Model calls still go to the chosen provider or gateway; this is not an on-premise release of Claude model weights. Anthropic says session histories stay in the customer's database, supported tool credentials can be held outside the sandbox, network hosts can be restricted, and activity records export through OpenTelemetry. Customers take responsibility for maintaining and securing their deployment.

Analysis. Hosting the agent control layer gives companies more control over operational records and credentials without removing the external inference boundary. That distinction matters for compliance claims: a locally hosted agent service is not a wholly local AI system. The trade is more infrastructure ownership and maintenance in exchange for greater control over unattended work.

https://claude.com/it/resources/articles/self-hosted-claude-managed-agents
https://www.helpnetsecurity.com/2026/11/12/anthropic-self-hosted-claude-managed-agents/

Scope and sourcing notes

Ten distinct underlying developments. Helsing unveiled Clarity on 10-11 November and appears here for the first time; earlier editions did not cover it. Runway's Praxis-1 research preview dates to September; today's news is the open-weight release, and its benchmark scores are company-reported against other vendors' published figures, not third-party replications. Qwen4-Max-Preview is a fixed two-week, Token Plan-only preview with no published benchmarks. Qwen Code 1.0 is a GitHub-marked pre-release, and AWS's agent-authorship account is the company's own. Google's Decisions API latency and cost figures are vendor numbers. Helsing and Databricks product claims are attributed in place. Excluded as stale or duplicate: GitHub Agentic Workflows GA (generally available since 28 October), NVIDIA Nemotron 3.5 write-ups (October releases covered earlier), Baidu's ERNIE 5.5 Flash Preview API access (derivative of the 7 November open-base release), follow-on Anthropic IPO valuation chatter after yesterday's dated clarification, and coverage of DeepMind's AlphaProof Code Signal verification (11 November edition). Also weighed and dropped: Japan's AI penalties proposal, the EU regulatory sandbox summary, Gemini Live on CarPlay, Google Labs Tabletop, Tether's BotPsy on-device agent, n8n's Nod and OpenAI's caregiver workshops. No public X post is used as a Firsthand signal today. Publication update, 13 November: Australia's independent-testing proposal repeated the 10 November announcement and has been replaced by Anthropic's 12 November self-hosted Managed Agents private beta. The Word ledger remains unchanged; this post includes the later verified replacement.

Bring agentic AI to your classroom

TechSambad's AI Learning Pilot puts agentic AI in front of your students for one module, not a semester: a live 90-minute taught session inside your timetable, an AI-built pre-class reading brief mapped to your module, a question bank with grading rubrics, a student takeaway pack, and a written outcome report. Run it once and judge it on what your students can do afterwards - Rs 2,000 + GST for the college, no obligation.

https://www.techsambad.com/p/ai-learning-pilot.html?m=1