Joseph Varga

Founder & CEO @forgeorbital. The system questioned cannot be its own proof. Forge is now available in ChatGPT:

[Joseph Varga @joeyvarga_]

AI is starting to act inside companies. Who owns the call?

[Joseph Varga @joeyvarga_]

@layerlens_ai I'd add an in-flight case to that test: a write starts before the stop and finishes after it. Can the record distinguish that from a new action started after the stop?

[Joseph Varga @joeyvarga_]

Accountability needs a stop button that actually stops the work. If an agent hands a task to three others, a human override should reach all three, not just the first chat. How do you test that the stop reached the whole chain?

[Joseph Varga @joeyvarga_]

@_philschmid Fast branching is useful. A confidence score still needs a rule for when to stop. Can the application preserve the threshold, input and model version behind each route to a human?

[Joseph Varga @joeyvarga_]

@NVIDIAAI @nvidia Congrats to the builders. My favorite demo question: can you replay the failed attempt too, and show which check caught it before the agent acted?

[Joseph Varga @joeyvarga_]

Thanks Brittany for another great article.

[Forge Orbital Inc @forgeorbital]

What happened is only half the audit trail. Can yours explain why? Read Brittany Istenes' new Industry Pulse article: https://t.co/eOrkpPiijb

[Joseph Varga @joeyvarga_]

@GoogleLabs The group chat finally gets a plan. Can everyone see which parts came from a person and which Tabletop generated, especially when someone changes the budget?

[Joseph Varga @joeyvarga_]

Accountability gets tested when the task changes. An agent approved to draft a refund should not treat that as approval to issue it. What change in scope forces a fresh approval in your workflow?

[Joseph Varga @joeyvarga_]

@levie Read access finds the contract. Write access changes the deal. Can the review record show what the agent changed, which instruction allowed it, and where a person had to approve?

[Joseph Varga @joeyvarga_]

@TechCrunch "We should do that" in a meeting is not always an instruction. How should an AI teammate tell brainstorming from an approved task, and keep the difference in the record?

[Joseph Varga @joeyvarga_]

@hedra_labs Recursive improvement makes the action boundary matter. When a training or inference job moves from proposal to run, what record shows the approved inputs, the rule that allowed it, and who could stop it?

[Joseph Varga @joeyvarga_]

@OpenAI Business agents make ownership concrete. Someone has to keep the source information current and decide when an answer needs a person. How will customers see that handoff?

[Joseph Varga @joeyvarga_]

Accountability for AI agents starts before the action: the request, the rule, and the evidence have to be in one place. If one is missing, the default should be hold, not a postmortem. What is your default?

[Joseph Varga @joeyvarga_]

@unusual_whales The number matters less than the control point. If agents quote or execute bond trades, which decisions need a person’s sign-off, and who keeps the record when one is skipped?

[Joseph Varga @joeyvarga_]

@OpenAI A central stop switch is the right instinct. The hard part is deciding which signals trip it and who can clear a stop. How are those decisions recorded?

[Joseph Varga @joeyvarga_]

@emollick The hands-off part is the real stress test. At 25,000 agents, who reviews the exceptions, and what record shows why they were allowed?

[Joseph Varga @joeyvarga_]

@AIatMeta Cheap enough to run all day also means cheap enough to run unchecked. For a small model making thousands of tool calls, who sets the limit on what it can do?

[Joseph Varga @joeyvarga_]

Before an agent acts on a Monday task, I would check three things: what it is allowed to touch, who approved that scope, and what happens when it hits the edge. Which of the three does your team have written down?

[Joseph Varga @joeyvarga_]

@NVIDIAAI @Hyundai @AIMSafety_ @LG_AI_Research Secure by default is the right place to start for enterprise agents. Next question is who decided a given action was allowed. Do teams keep that decision for each action, or just the sandbox rules?

[Joseph Varga @joeyvarga_]

@NousResearch Agreed, and the quiet part is that you often find out after the output changes. Is a visible change log for the model and code enough, or should the agent have to ask first?

[Joseph Varga @joeyvarga_]

@github @code The handoff from agent to person is a good moment to check the work. When the task lands in the IDE, can you see what the agent was allowed to do, not just what it changed?