A traceable decision trail increases human confidence
You don't need to show all the logs and math behind a single number, but having a trace to follow back a decision increases confidence for humans quite a bit.
40,602 results · Most relevant
You don't need to show all the logs and math behind a single number, but having a trace to follow back a decision increases confidence for humans quite a bit.
When building a harness, show the receipts — where something came from and why it is correct — establishing a causality relationship; this builds trust in the agent's output.
Build more feedback loops so you don't have blind spots, but without relegating any accountability. Build gates for validating things, staying accountable, and correcting course safely.
Your definitions of certain things will not exactly match the model's unless you have models trained for you. Go back and forth with your model to align definitions, because what the model knows is not necessarily what you think.
Handing over entirely the keys to the kingdom to something you don't necessarily trust to have your best interest in mind is a mistake.
Make sure all components of your harness work together as they should rather than based on hopes, dreams, and magic glue from an agent orchestrating things.
Figure out what you actually care about and what the functional requirements are for a functioning harness. Make sure the objectives of each step have an intent.
Prioritize trust-building mechanisms over just giving something that looks correct when building a harness.
Build the feedback mechanism on purpose and have some judgment over making sure there is some consequential decision-making on the fact that an input may change your whole analysis.
Build trust with your agent by making sure the information you care about about its implementation is in front of you. The intent of the UI was aggregation of everything: all steps showing where results come from, that it's running SQL, that it's tested.
Clark argues that any proof or formal safety guarantee holds only with respect to specific assumptions, and that the formal methods community has learned to state those assumptions explicitly or face disappointed users. He illustrates with the case of assuming another real-world actor won't act adversarially, which may not be valid.
ZocDoc now exposes the scheduling infrastructure behind its marketplace to payers, AI platforms, and consumer-audience companies, so patients can book doctor appointments from wherever they are in their journey. Kharraz describes the company's role as healthcare access infrastructure: it does not need every booking to happen on its own app, only that whatever door the patient knocks on actually opens. The stated goal is to connect every provider and make every 'front door' in healthcare work, letting payers, AI platforms, consumer platforms, review platforms, and search sites connect patients directly to doctors.
Reserve flows for ambitious, complex tasks; flows let you throw agents at a problem in an organized way rather than an unorganized way.
Letting a thousand agents run wild on the internet without safety mechanisms is unsafe; agents scanning the open internet can pick up instructions by accident from people or entities that are not yourself.
Starting a large multi-agent project and abandoning it without monitoring mechanisms risks losing visibility into intermediate states — what if the thing dies, and how do you see the intermediate states? Monitor telemetry during agent operation to determine if you need to take actions based on what the session is doing and whether you're going in the right direction.
Hormozi argues that the true losers are not what limit a business; it is the mediocre middle, whose slow, mediocre work clogs the pipes and forces high performers to rework it.
The lethal trifecta is a security concept coined by independent AI researcher Simon Willison describing the severe vulnerability AI agents get when they get simultaneous access to three different capabilities. Having any one of these is usually fine because your AI tool won't be very useful if it can't have access to any of these things. Having two together is risky and not something the speaker would recommend, especially in a regulated industry like banking or financial services. Even the lethal trifecta isn't a checklist to take off and then you're safe; it's more of a mental model for understanding how much risk your AI tooling is actually exposing you to.
The team found it important to keep the document structure — the section part matters both for retrieving the answer and for the agent to zoom out on the knowledge — and to give the agent multiple zoom levels, letting it decide how much context it needs.
The team found it important to test with real users early on, and to use those user interviews to create promptful evaluations that shape the product and verify whether a perceived improvement is actually an improvement.
It's fine for the LLM to know what a bank is, but knowledge about other payment schemes — what they charge or how they work — must not end up in answers.
Choose sources → to get a library of your own and a daily brief.