417 results · Most relevant
The hosts recount that basically every single conversation with an agent — whether coding or working on a deck — involves asking 'where did this number come from?', and it would be nice to just be able to show the paper trail.
The hosts recount that when you ask for a modification and give an example, the model will fixate on the example so hard that it will put it everywhere in your presentation.
Francesco tells attendees to expect something to break during the session and to call the presenters over to fix it with them, and to raise their hand if something is not working or they are confused so the presenters can float around the room.
Felippe invites attendees to type something they want to use a harness for in their own use case, such as planning a trip, a new investment in their company, or in their personal life.
The hosts suggest adding a little bit of a curveball to the agent by typing it in the chat, and you will see how some of those evidence lines or inputs that go into the rest of your process change.
Moshi wants to understand what problem a prospect is solving for and have a plan after the initial meeting. Due to the amount of inbound, they are hosting their first webinar because they can't always do one-to-ones. The current priority is meeting clients rather than hiring.
Anekha Sokhal states that the priority right now is to meet clients rather than hiring, because they receive continuous inbound from LinkedIn for hiring.
Anekha Sokhal describes significant noise in hiring: people shopping or spying in an undercover way, people legally not allowed to work, and visa complications. She has a strong opinion that $2,500 is too much and most of it is noise, arguing a better filter is needed on the hiring side. She also wants to help the interviewee on the hiring side.
A principle is something you need to interpret; don't leave people with a vague, ambitious thing like 'do evidence, it's awesome' without specifying what about evidence they need to take care of.
Francesco Galletta frames the goal: everything starts from earning trust, so that you can truly delegate the boring work to multiple agents, swarms, or factories.
Building taste to create fit-for-purpose harnesses for your use cases is part of the design. Being intentional with the design of your harnesses is an important part of the process.
Francesco Galletta describes a mindset shift: from throwing problems at black boxes hoping something sticks, to building systems that evolve with you. He also notes there is no single silver bullet to solve every problem you will face with agents, and that building is not the hard thing — the hard thing is knowing how you want to approach the problem; asking an AI to build you a million-dollar company without knowing the underlying principles leaves you relying on randomness.
Models are going to get better, but they still have to earn the trust — it's not 100% right.
endorsed
The point of the harness is to guide you and force you to do the right thing, and the principles are there to help you make the right decision every time.
Anekha Sokhal describes the reality of hiring for AI roles: one job posting drew 2,500 applicants in four days, most of it noise — people legally not allowed to work, visa issues, people 'shopping' or spying. She recommends applying a stronger filter because most of the per-candidate sourcing cost is noise, and treating visa status as a screening factor. She notes first-time founders typically don't know these hiring-market noise factors when they start.
Based on the speaker's own hiring experience; the 2,500-applicant figure is her report.
Anekha describes Bloomberg chat as a fantastic aspect of the platform, noting that traders would transact billions of dollars in chat using a standardized formula. Her own entry into NLP was parsing Bloomberg chat into a structured pricing sheet for her logs.
At Monzo, there's an internal contribution mechanism for agent extensions where multiple teams can share what works — plugins, skills, and so on. When reviewing contributed skill instructions, Monzo frequently sees things trying to cover for cases like 'if the user tries to do this, tell the user back that this skill does not support that' — attempts to cover for many things that determinism (back-end code) covers out of the box.
When presenting AI-built tools to senior leadership for financial decisions, there is an expectation that the output will be 100% correct.
described
Agents will probably perform your instructions 99% of the time, but the 1% edge case is what matters when distributing to many users.
speaker's estimate, not measured
Other implementations of financial analysis in the industry provide traceability down to the dataset: you know which connector or supplier gave you the information, e.g., that Apple had a certain price at a certain point of the day.
No published protocol matches this selection.
The creators behind most of what cleared.
344 cleared insights match
71 cleared insights match
Choose sources → to get a library of your own and a daily brief.
