Can an AI agent legally run my business?

Not as things stand, and the idea is contested. Ars Technica reports that a Delaware proposal would create a corporate form operated by AI agents, and that the chair of the House Oversight Committee has written to Delaware's governor arguing it would help owners evade responsibility for the AI's actions. Whatever happens there, the customer still holds you accountable for what your systems do.

Should I ask AI whether my business idea is original?

Use it as one input, not the verdict. In an Allen AI benchmark, GPT-5.4's accuracy at picking the more novel of two ideas fell from 93.5% to 40.9% when the prompt wording changed, and it reversed more than half its answers. Ask the same question several ways, and check the answer against real competitors.

This Week in AI: No Human at the Helm, Fickle AI Judges

The week’s lesson is that someone still has to own the outcome. A Delaware plan to let AI agents run whole companies drew a sharp letter from Congress, Amazon told its engineers that AI not helping is worth reporting, and a benchmark showed AI judges reversing their verdicts when a prompt changed by a few words. For a solo operator, the useful question is not what AI can do. It is what you are still on the hook for.

Delaware weighs companies run by AI agents

Ars Technica reports that Delaware is working on a new corporate form operated by AI agents, and that Rep. James Comer, who chairs the House Oversight Committee, wrote to Governor Matt Meyer arguing it would help human owners evade legal responsibility for harmful actions taken by the AI. The committee is asking for documents about the effort. I have only the article’s summary, not the proposal itself.

What this means for you: the legal debate will take time, but your customers decide now. If an agent sends the wrong invoice or makes a promise you cannot keep, the person on the other end blames you. This is The Intent-Execution Gap: an agent executes what you said, not always what you meant, and the gap is yours to close with clear limits on what it may do without you. The test I use is blunt. Would this keep working if I were away for a week, and would I stand behind every action it took?

Amazon: AI not helping is a finding

Simon Willison quoted guidance from Amazon engineering leadership this week: if current AI tools do not help your work, “that is a finding, not a failing.” Engineers are asked to write down why and bring it to their manager, and the guidance calls that explanation good judgment.

Solo operators rarely get that permission, because every feed says you are behind if AI is not doing everything. The Amplifier Paradox is the reason to take Amazon’s line seriously: AI can raise your output while quietly weakening the skill that made the output worth paying for. Keep a short list of tasks where AI does not help you, and why. That list is part of your expertise.

AI novelty judges flip on a few words

Allen AI published a benchmark on whether AI can judge which research idea is more novel, built from ICLR 2026 peer reviews. The striking result: when the prompt changed from one that said reviewers found one idea novel to a plain “which idea is novel,” GPT-5.4’s accuracy on pairs that included AI-generated ideas fell from 93.5% to 40.9%, and it changed its verdict on more than half the pairs. Claude Opus 4.6 still shifted 21 points on human-written ideas.

If you ask AI whether your offer, course or positioning is original, this is the warning. The answer depends on how you ask. The Belief Offloading Spectrum separates AI that informs your judgment from AI that forms it. Ask the same question three ways, and treat disagreement between the answers as the real signal.

Frontier models you cannot use yet

Google announced Gemini 4 Carbon on October 28, two weeks after its previous frontier model, and is giving it first to cyber defenders, with no release date for businesses. OpenAI shipped GPT-6.1 Terra days earlier. Even game-making moved: Simon Willison argues that October 2026 is an inflection point for vibe-coded games, and says the bottleneck is now finding ideas worth building rather than technical skill.

Two frontier announcements in one week is a reminder not to tie your business to any one model. The leaderboard will reorder again next month. What carries over is what you own: your instructions, your examples, your record of corrections. That is the argument in Marketing-Grade Decays. Engineering-Grade Compounds.: a system that learns your business keeps its value when you swap the model underneath it.

The throughline

A state debating agent-run companies, an employer asking for honest reports of where AI fails, and judges that change their minds with the wording all point the same way. The model is the part you rent. The judgment about what it may do, where it helps and when to trust it stays with you, and so does the responsibility.

If you want a system that learns your business instead of renting a stateless tool, start at curiochat.ai/solopreneur/.

This Week in AI: Companies With No Human at the Helm, and AI Judges That Flip on a Few Words

Delaware weighs companies run by AI agents

Amazon: AI not helping is a finding

AI novelty judges flip on a few words

Frontier models you cannot use yet

The throughline

Related