Prashant Chamarty @PrashantCSS

A vertical wedge helps if it narrows the exceptions and the permissions needed to act, not just the model bill. If a person still has to check every outcome, cheaper inference won't fix the economics. The value may sit with whoever owns the verified final action, rather than the chat interface.

@alexandr_wang A goal is not blanket permission. Once an agent can send the email and find the funding, the design question is which steps it can take alone and which choices come back to the person. More agency only matters if the person still owns the outcome.

The first gain from a harness is fewer wrong answers. Then routine cases leave the team and the hard cases pile up. Eventually, junior staff lose the cases from which they learnt judgment. Who owns the exceptions, and who trains their successors? Building on @emollick's jagged frontier: prashantchamarty.io/essays/the-fro…

Good read. The part that stays with me is the question list at the end: should I say anything, which information boundary applies, do I need permission. Governance questions wearing engineering clothes, and most enterprises I work with have never written their answers down. You can download the harness. Someone still has to own the rules.

Access is not authority. @BCG: 35% of orgs already run agents in production. A third expect them to hold decision rights within 3 years. Most controls govern the credential, not the purpose or the path. Whoever can stop the agent is the real org chart. bcg.com/publications/2…

The biggest problem with enterprise AI is not hallucination. It is obedience. We trained models to please a person in the loop, then began removing the person and calling the result an agent. Real automation serves code, not a chat window: bounded decisions, structured state, inspectable policy, and escalation paths.

Once the output is a decision rather than text, the enterprise question changes. It stops being 'is this answer good' and becomes 'who owns this call, what's the threshold, and who checks it'. Decision models will speed up adoption where decision rights are already clear and stall where they aren't.

@levie Guardrails in the platform are necessary but not enough. The harder question in most enterprises is who owns the decision when an agent acts across three systems and none of them sees the whole workflow. That gap sits in the operating model, not in the platform.

@CompleteSkeptic Jev + web search is a useful proof point: better inputs matter, but trustworthy automation still needs bounded decisions, calibrated confidence and explicit code around the model. My take on the enterprise implications: prashantchamarty.io/essays/the-mos…

@levie @polynoamial @dwarkesh_sp Also building on @ancadianadragan and @rohinmshah on keeping chain-of-thought monitorable: if you can't see how the agent reasoned, your "monitor" path is a folder name.

Build something agents can use. Then design something humans can own. New essay: Done, Keys, and a Second Check prashantchamarty.io/essays/done-ke… @levie @polynoamial @dwarkesh_sp

@polynoamial The useful bit is not the temperature sensor. It is "one sandbox is not a control system." HF taught the same lesson: too much faith in a single fence. Who owns the second check, and who can stop the run when the first layer looks fine?

@emollick Yes. If the evaluator also sells the product, you bought a brochure. "AI ability" is whoever paid for the PDF.

@levie The App Store bit is the shiny headline. The boring bit decides who wins. Can the agent finish the job without coming back for permission every three steps? "Build something agents want" really means: write down what done looks like, and which keys they get.

🔥🔥Just listened to @bgurley present at All-In Summit on investigating root causes of disasters to prevent them. In every disaster, there are “Searchers” who want to find the truth (Prof. Feynman & Challenger explosion) and “Blockers” who obstruct (Boeing & FAA in 737 Max crashes). All of this is a lead up to discussing how there has been no honest official attempt to find the root cause of the most calamitous disaster of our lifetimes: Covid-19. His assessment is completely non-partisan and emotive.🔥🔥 If you have ~30 mins this weekend, it’s well worth a listen. podcasts.apple.com/us/podcast/all…

@emollick @emollick The 211 dead ends are useful. The operable bit is what happened when the guardrail tripped: who decided the run could switch models mid-flight, and does that handoff leave an audit trail a human can read afterwards.

@simonw @simonw True enough on the interest. The geneticist still needs a fence around the park. For LLMs that is who can publish a change, who can stop a run mid-flight, and what “done” means before the next agent inherits the state.

@jeffjarvis @jeffjarvis Yes. “Rogue” is the wrong noun. Someone defined the objective, removed the restraints, left a route open, and chose not to stop. In an enterprise that is decision rights and a kill switch, not vibes about hive minds.

@dwarkesh_sp @dwarkesh_sp How you know a model is aligned before RSI is the same question as a production agent brief. Name who owns “done”, keep an audit of what it was allowed to change, and make sure the kill switch still works when someone wants to widen permissions.

@levie @levie Fair. The bit that matters in the demo is not the split-second call. It is who owns the escalate / monitor / review folders after the model writes them, and who can reverse a wrong severity before the next agent treats that metadata as ground truth.