AI Daily Digest November 12, 2026: AWS Emulator Written Entirely by AI Agents, Google Scores Decisions in One Pass

Two product stories and five funding rounds set the tone today. AWS shipped a tool whose entire codebase was written by autonomous agents, Google turned closed-set decision making into a priced API, and investors pushed hundreds of millions into chips, welding robots and even orbital mirrors. Here is what happened over the past 24 hours and why each item matters.

1. AWS shipped an emulator written entirely by AI agents

AWS released AWS Emulator, a free open source tool that replicates more than 200 cloud services locally so developers can test code without touching the real cloud. The remarkable part is not the feature list but the origin story: AWS states that thousands of autonomous AI agents wrote every line of the code, and that the team did not write or review it. The announcement was published on Amazon's official AI news page on November 11.

The project was conceived by Matt Goldberg, a senior principal engineer at AWS. His inspiration was Alf, a Paxos based locking service built more than two decades ago that quietly underpins Amazon S3 and DynamoDB. Using AWS's public Smithy API models as the formal specification and the live production cloud as ground truth, he prompted AWS Kiro to reproduce Amazon SQS. The agent produced a functional, compliant replica in 15 minutes.

Scaling from one service to hundreds took an assembly line of dozens of agents running on Amazon Bedrock AgentCore, organized into three groups. One group generates thousands of end to end integration tests from the API models and verifies them against the live cloud. A second runs those suites against the local emulator, files structured bug reports on GitHub when behavior diverges, and a third writes patches. Before any patch lands, it passes an automated merge gate supervised by adversarial security agents that probe for vulnerabilities.

The tool ships as a single Rust binary and a Docker container. Developers point AWS_ENDPOINT_URL at localhost port 4566 and standard CLI and SDK calls resolve locally, no code changes required. It is released under CC0 1.0 with the full raw commit history attached, and the public issue repository carries a label called things-that-do-not-work to track remaining edge cases. The economics are the point: a coding agent running write, compile and test loops can trigger hundreds of provisioning cycles per hour, remote round trips cost 150 to 600 milliseconds versus under 5 locally, and AWS claims local validation accelerates agent iteration more than 20x while eliminating cloud testing bills. Swami Sivasubramanian, AWS vice president of agentic AI, framed the shift plainly: engineers become systems orchestrators whose job is defining boundaries, specifications and adversarial test suites. For teams running agents in production, a first party emulator that regenerates itself whenever the Smithy models change is a meaningful piece of plumbing.

— AWS · AINewsToday

🔗 AWS · AINewsToday

2. Google's Decisions API scores 1,000 questions in a single pass

Google introduced the Gemini Decisions API on November 11 along with a purpose built model, Gemini Decision 1, available through the Gemini API in Google AI Studio. Instead of generating answer text token by token, the model scores candidate options in a single forward pass and returns probabilities plus a confidence score from 0 to 1.

One request carries a shared multimodal input across text, images, audio, video and PDFs of up to 256K tokens, and evaluates up to 1,000 independent questions against it. Three question types are supported: a yes or no probability, multiple choice with 2 to 255 options returning normalized probabilities for every candidate, and ordered rubric scoring that returns a continuous weighted score with per level probabilities. Every answer includes that confidence value, so application code can route ambiguous cases to a fallback or a human reviewer instead of guessing.

Google reports end to end latency of 150 to 250 milliseconds with median model compute of 50 to 100 milliseconds for contexts up to 16K tokens, and the numbers stay nearly flat as you scale from 1 to 100 questions per request. Pricing is USD 0.10 per 1 million input tokens with zero output token billing, which the company says is roughly 5x cheaper across a shared benchmark suite than Gemini 3.5 Flash-Lite running through the Interactions API. A no cost tier exists in AI Studio.

Internally Google already uses the endpoint for dynamic model routing in Notebook, cutting daily query latency by up to 40 percent, plus wireframing in Stitch, emoji selection from more than 200 options, and a safety layer for MCP tools in Search. Integrations are rolling out with OpenRouter, LangChain, Vercel AI SDK, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode. The bigger picture is that OpenAI shipped its own Decisions API in late October, as we covered in our November 6 issue. When two frontier labs independently turn closed set judgments into a priced endpoint, routing, ranking and policy checks stop being prompt engineering chores and become infrastructure calls with predictable cost and latency.

— Google · Sundar Pichai

🔗 Google Blog · Sundar Pichai

3. Mercor doubles to $20B with a $750M Series D

Mercor announced a $750 million Series D on November 10 at a $20 billion valuation, doubling the $10 billion from its $350 million Series C in October 2025. The round includes an employee tender alongside Benchmark, Felicis, Fundomo, funds managed by affiliates of Apollo, G Squared, General Catalyst, Glade Brook Capital, NVIDIA, Salesforce, Sapphire Ventures, StepStone Group, XTX Ventures and Zoom. No lead investor was named.

The company made its name supplying human expertise to frontier labs: more than five million paid experts train and evaluate models on professional work, drawing on what Mercor calls more than 800 distinct US occupations whose workflows have yet to be evaluated against any model. The new pitch extends that same infrastructure to enterprises. Mercor's playbook starts from a capable open model, builds custom evals from a company's actual workflows, post trains to cut costs, and then trains models for that company's customers. Earlier this year it acquired Deeptune, whose team spent two years recreating hundreds of enterprise applications as reinforcement learning environments where agents can practice professional workflows.

The financial contours are striking. Mercor said in July that annualized revenue had passed $2 billion, and the company spends three times as much on inference as on employee salaries. CEO Brendan Foody argues that if that ratio holds as AI spreads through the economy, it points toward something closer to double digit GDP growth over the next five years. The round lands in a competitive market against Scale AI and Surge AI, and Mercor has opened offices in New York and London this year to chase enterprise buyers. The signal for builders: the buyer of evaluation and post training infrastructure is expanding from a handful of frontier labs to every company that wants to own its intelligence.

— Mercor · TechFundingNews

🔗 Mercor · TechFundingNews

4. Britain backs Fractile with a £500M+ round at £5B+

The UK government announced on November 11 that its Sovereign AI Venture Fund and the British Business Bank are backing Fractile, an AI chip startup based in London and Bristol, as part of a Series C round worth more than £500 million that values the company above £5 billion. Founder and CEO Walter Goodwin put the raise at $720 million, and Reuters reported a valuation near $7 billion. Thrive Capital, Founders Fund, Lightspeed and Redpoint led, with CapitalG, Shopify, Balderton, Dragoneer and BlackRock also participating.

Fractile is attacking the inference bottleneck. Its architecture keeps data directly on the chip instead of shuttling it between the processor and separate memory modules, which the company identifies as the dominant cost and speed constraint when running large models. The claimed numbers are aggressive: LLM inference up to 100 times faster than existing hardware with operating costs cut by 90 percent. Those are company figures, not independent benchmarks, and the proof will arrive only when the first chips reach data centers, which Fractile says will happen next year with first chips and systems shipping in 2027.

The pace of the revaluation is notable on its own. Fractile raised a $220 million Series B in May at around £820 million, and six months later private capital plus the British state are pricing it above £5 billion. For the government this is industrial policy: AI Minister Kanishka Narayan framed the investment as keeping the value of British breakthroughs at home rather than repeating the pattern where British science becomes someone else's industry. The autumn Budget added £140 million for Sovereign AI, split between fund investments and an R&D procurement scheme, and the initiative is extending into robotics, autonomous systems and advanced manufacturing. If you buy AI compute, the number to watch is the deployment date, not the press release.

— GOV.UK · Reuters

🔗 GOV.UK · Tech.eu

5. Path Robotics raises $300M to scale factory physical AI

Path Robotics announced $300 million in Series E financing on November 10 at a $2.45 billion post money valuation, led by Georgian. Existing investors Addition, Drive Capital, IronArc Ventures, Tiger Global and Yamaha Motor Ventures returned, joined by new backers D. E. Shaw Ventures and NAVER Ventures, and Georgian partner Emily Walsh joins the board. Total funding now stands at roughly $610 million.

The Columbus, Ohio company builds what it calls physical AI for manufacturing, and it started with welding, one of the hardest factory jobs to automate. Its Obsidian model runs on real world welding data and a learned weld world model: the robot identifies weld joints, plans and executes welds, and adapts in real time to the heat induced distortion and part variations that defeat pre programmed industrial robots. In April the company introduced Rove, a mobile welding system mounted on a quadruped platform aimed at shipbuilding and large fabrication work.

Commercial traction justified the round. Annual recurring revenue grew 17 fold over the past 18 months, the workforce grew 97 percent year over year with a target of 400 employees by year end, and customers include HII, the largest US shipbuilder, plus Lockheed Martin and Saronic alongside dozens of mid sized manufacturers. Earlier this year Path signed a $600 million performance based production agreement with HII to accelerate deployment across US Navy shipbuilding programs. CEO Andy Lonsberry set the ambition plainly: the goal is not to automate one step but to automate the entire factory, which means extending the models to every task before and after welding and to new robotic form factors. For anyone tracking physical AI, this is the funding proof that welding robots moved from demo videos to production lines with defense budgets behind them.

— Path Robotics · Bloomberg

🔗 Path Robotics · Bloomberg

6. Plurai raises $13M and opens agent simulation to everyone

Plurai announced $13 million in seed funding on November 11 along with the general availability of its platform, now open to all agent builders with a free tier. The round was led by Team8 with participation from U&I, Velocity and Mercer, backing from more than 20 angels including CyberArk founder Udi Mokady, and Microsoft 365 Copilot CTO Eyal Carmi joining the advisory board.

The product addresses a failure mode every agent team knows: internal evaluations pass, production fails. Plurai builds a digital twin of an agent and its world from the customer's own knowledge base, tools and policies, then runs thousands of realistic and adversarial simulated users against it in a sandbox. That surface exposes policy violations, hallucinations, incorrect actions, broken tool calls and sensitive data leaks before real customers ever hit them. The platform then generates fixes and uses the simulation data to train small language models tailored to each agent, a synthetic data method called BARRED that was accepted at ICML 2026.

In production those SLMs play two roles. They can replace frontier models to run the agent at a fraction of the cost, and they act as always on evaluators and guardrails checking every action against company policy, with new failures feeding back into the simulation. The reported numbers: design partners cut development cycles from months to hours, failures reaching users dropped 82 percent, and the cost of running reliable agents fell 97 percent. In joint work with NVIDIA and Nebius for a global enterprise customer, Plurai post trained models outperformed OpenAI and Anthropic frontier systems on accuracy at roughly 40x and 44x lower cost. Founders Ilan Kadar and Elad Levi come from Nexar and the autonomous driving world, where Waymo earned trust with 20 billion simulated miles before touching real roads. Their bet is that agents need the same treatment, and that simulation is how you get it.

— Plurai · PR Newswire

🔗 Plurai

7. Reflect Orbital raises $73.5M to sell sunlight on demand

Reflect Orbital raised $73.5 million on November 10, bringing total funding past $100 million. Sequoia Capital, Lux Capital and Starship Ventures returned, joined by Buckley Ventures, Convective Capital, The Uncommon Group and Salesforce Ventures. The company, founded by former SpaceX and Zipline engineer Ben Nowack together with Tristan Semmelhack, builds satellites that redirect sunlight to specific locations on the ground.

The mechanics: each satellite launches folded and opens a square mirror roughly 60 feet across, steering sunlight onto a spot about five kilometers wide at approximately the brightness of a full moon. A single satellite illuminates its target for about four minutes per pass, so a commercial service needs a constellation. The company says it builds a reflector quadrant 28 times thinner than a human hair every 3.5 days and assembles a 12,000 part satellite from scratch every two months. The pitch covers extending solar farm output after sunset, lighting disaster zones for search and rescue crews at night, and illuminating remote industrial sites.

The first demonstration satellite, Earendil 1, holds an FCC license and is planned for launch in the second quarter of 2027. On October 14 the company bought two dedicated launches from India's Skyroot Aerospace, a Vikram 1 and the larger Vikram 1U. The idea remains polarizing: the American Astronomical Society said it was dismayed by the license, while the FCC judged that a single satellite was unlikely to raise astronomy or environmental concerns and declined to rule on the larger constellation. Reflect Orbital says it maintains strict exclusion zones for astronomy and sensitive environments, and its roadmap stretches from 2 satellites in 2026 to 50,000 or more by 2035. Whether sunlight ever becomes a billable utility is a question for 2027, but investors just put fresh capital behind the attempt, and the FCC process has already made this the most scrutinized mirror in orbit.

— Reflect Orbital · SpaceNews

🔗 Reflect Orbital · SpaceNews

KD Agentic · AI Daily Digest