417 results · Most relevant
The hosts argue that the trust customers have in your product, your agent, and the output is what will drive usage, adoption, improvement, and so on, through the loops mentioned — tracing, evals, and so on.
In the demo, evidence is highlighted and users can click to see the areas they care about. The flow of information through the harness is visible, including which parts were proposed by the assistant, so users can 'show the receipts' of where a result came from.
Doing financial analysis in a harness without these principles will probably have some of the pitfalls that Francesco talks through in each of the steps — so it makes sense to build a thing to avoid these pitfalls seen over and over and over again. The recipe is specifically good for financial harnesses because there's a lot at stake — you want a lot of feedback loops, and there are very specific things about those use cases. The recipe is specifically geared towards financial use cases and can be extended to different levels of it, but everything hinges on the fact that it's a trust system — it may not be as useful if your use cases are not necessarily like that.
The branching harness gives a means to be accountable for what goes into the system for financial analysis and what doesn't. In the wedding example, you can hover over the chip that guessed how many guests there are, add it to the chat, and say 'oh, actually it's not this number. We will get no fewer than 150 guests.' Correcting an input gets the harness to rethink the plan, see the consequences of that input on other things, and suggest either a modification of your plan or just change the input.
Felippe explains that when tests are passing, you only need to check that everything is passing and you will have a lot of confidence because of it.
Premises will change, source data will change, context will change — having the harness able to adapt to that means recognizing that change will be a thing, rather than having to prompt again every time and go back to the process.
Anekha Sokhal describes Moshi's product goal: providing real-time insights for financial traders, starting with commodity, energy, and macro markets — the program she wishes she had when she was trading. She frames the value as giving traders answers in 30 seconds that would typically take days, and notes Moshi processes data at a scale where in the four seconds it takes to speak a sentence, it has learned 300 new things.
Product claims from the founder; the '300 new things in four seconds' and '30 seconds vs days' framings are her own marketing-style figures without independent verification.
Some conversations don't change anything or aren't material to your process, and you will know better; open a challenge branch and click dismiss as a deliberate design choice.
endorsed advice
Francesco recommends anchoring the principles again because they build on top of each other in a sequence that makes logical sense.
If you're trying to use something and not getting the right results, try adding a couple of skills, commands, or extra things before building a harness.
endorsed advice
Work with SMEs, the client team, or whoever it is to understand what would be the best set of things that agents use to get to the best answer.
Felippe recommends that when building financial harnesses, you can guide users in how much they want to be involved in the process, in the same way Claude Code does.
For pick-your-own-adventure, type a financial problem you want to use the harness for, such as planning a trip or a new investment in your company or personal life. Note the added risk that the presenters may not be able to help you if you get stuck.
When building a harness, ensure a feedback mechanism so you can verify from what the model thinks it knows all the way to what you think you know.
endorsed advice
Anekha describes that divergent sources (Reuters, Bloomberg, WSJ disagreeing) are themselves a point of data because it may indicate an opportunity where no one knows the answer.
Anekha describes that Moshi has an internally built recommendation system for traders, and argues that recommendation systems seem easy but are arguably the most important system in the entire company. She connects two simultaneous information points either via a human or by mimicking a human with a recommendation system.
The thin slice and the quick instinctive look also come from the trace, because you understand the thought process that the agent went through, and not only the output. It's kind of satisfactory when the thing just goes into it. The branch comes back — and it creates a trace so you understand the thought process the agent went through, not only the output.
Don't over-engineer it — it's easier if all the components of your harness just work together as they should and not based on hopes and dreams and magic glue from an agent orchestrating things together.
The harness UI's intent was aggregation of everything — all the steps in the framework should signal trustworthiness: 'I know where this is coming from', 'this is running SQL', 'this is tested' — an experience to immerse into that feels trustworthy. There won't be a single thing that works to earn the user's trust at any point. For some users, verification will be more important; others are more skeptical and want to see the actual calculations and numbers — the process guides and covers a multitude of possibilities.
Thin slicing means having thin slices of your anecdotal experience so you know what feels right or wrong based on domain experience; a trace of the agent's thought process enables this quick instinctive look, not just output review.
described concept
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.
