AI news for Thursday, November 12, 2026: United bans AI agents from booking, Google's Gemini Decisions API

One airline locked agents out of checkout while Google and AWS showed how to make agents cheaper and easier to trust. Build the guardrails before the feature.

Agents Policy3 min
United bans AI agents from booking, even with your OK
United's contract of carriage, revised Nov. 11, bars AI agents from making, changing or servicing tickets, even with a human's approval, and lets it cancel those bookings. Agentic shopping just hit its first big locked door.
Why it matters If you build or use an agent that books travel, a United booking it makes is now a contract breach United can cancel. Keep agents on search and comparison, and have a person finish the purchase on United's own site.
SRCView from the Wing
Models Products3 min
Google's Gemini Decisions API bills $0 for output
Google's Gemini Decision 1 answers up to 1,000 yes/no, multiple-choice or rubric questions about one input in a single pass, at $0.10 per million input tokens and nothing for output. Routing and moderation just got a cheap, dedicated tool.
Why it matters If your app spends frontier-model tokens deciding which model to call, what to block or whether a draft passes, this does that job for a fraction of the cost, with a confidence score your code can act on. Replay a week of those calls through it before you renew anything.
SRCGoogle
Agents Open source3 min
AWS releases an emulator no human wrote or reviewed
AWS released AWS Emulator on Nov. 11, a CC0-licensed local stand-in for AWS services that it says autonomous agents built while its team wrote and reviewed none of the code. The merge gate the humans built is the part worth copying.
Why it matters If you run coding agents, AWS just published the setup that made their output trustworthy enough to ship: tests checked against the real service, separate finder and fixer agents, and a strict merge gate. Copy that structure before you trust agents with code nobody reads.
SRCAmazon
Safety Research3 min
DeepMind's AlphaProof Code verifies Signal in four weeks
Google DeepMind says AlphaProof Code's Gemini-powered agents wrote over six million lines of Lean proving key security properties of Signal's protocol code in four weeks, work it says took expert teams years. Cheap proofs change what secure software costs.
Why it matters Formal verification was held back by how few people can write proofs. If agents write them and humans only set the goals and check about 1% of the output, a machine-checked proof becomes something you can ask your security vendors for, not a research luxury.
SRCGoogle DeepMind
Hardware Business3 min
UK chip startup Fractile raises £540M at over £5bn
Fractile, a UK maker of AI inference chips, raised a Series C of around £540 million, and the UK government puts its value above £5 billion, up from about £820 million in May. Investors are pricing in cheaper inference before the first chip ships.
Why it matters Nothing changes on your token bill this quarter: Fractile's chips aren't deployed and its speed and cost claims are its own. What changes is pressure on compute prices. If you're signing multi-year compute or API commitments, keep the terms short while challengers like this reach silicon.
SRCTech.eu

ShareX LinkedIn Hacker News Email

Prompt of the day

Find the AI tasks that are really yes/no decisions

› I use AI for these recurring tasks: [list 5 to 10 tasks you run with AI each week, one per line].

For each task, tell me whether it is really a closed-set decision (yes or no, pick one from a fixed list, or score on a fixed rubric) or genuinely open-ended (it needs written output).

For every closed-set task:
1. Rewrite it as one precise question with the exact answer options or rubric levels.
2. Write the criteria that decide each answer, in plain language, so two people would give the same answer.
3. Say how confident the answer must be before I act on it without a human check, and what should happen below that line.
4. Give two tricky example inputs that would test the criteria.

For every open-ended task, say in one line what I would have to fix in advance to turn part of it into a closed-set decision.

End with a table: task, type, the question, the options, the fallback when confidence is low.

Open in Magai

Works in any AI chat. In Magai, run it against several models at once.

Get every edition

This one arrived in inboxes at 6am ET.

United bans AI agents from booking, even with your OK

Google's Gemini Decisions API bills $0 for output

AWS releases an emulator no human wrote or reviewed

DeepMind's AlphaProof Code verifies Signal in four weeks

UK chip startup Fractile raises £540M at over £5bn