417 results · Most relevant
Anekha describes that the big thing right now is getting software to actually be something people will use, framing it as a bit of chicken and egg between UI and recommendation systems. She says the company has been trying internally to understand how UI relates to a recommendation system.
If you rely on a gigantic spec for something like Francesco's wedding, things will change and evolve, and going back to the ever-changing spec means relying on the non-determinism of the agent to interpret it. The harness essentially breaks down what would be the spec into opinionated sections. Under the hood, in the 'my' folder, everything is saved — think of it as the spec behind the harness. The same principle applies in spec-driven development for coding agents: something is the source of definition for the thing being built; the difference is that the harness surfaces that through opinionated workflows rather than a single document.
A stakeholder from a separate team flagged a dashboard number as wrong because they instinctively knew from information elsewhere that it couldn't be right.
anecdotal
When a challenge branch result comes back, the agent joins back into the trunk and changes the plan, so there is a consequence.
endorsed
If you want to use AI just to produce social media content — for example, flooding Instagram and TikTok with AI-generated videos — there are steps in this recipe you probably don't care that much about, like evidence gathering, verification, or accountability.
Francesco Galletta advises focusing on feeling how trust in a well-harnessed agent works rather than on how quickly the UI responds or how slick it is.
When building a harness for a specific type of use case, it's different from the coding assistant world: with coding assistants the code is a means to an end — you're building something that performs an action, a product, an app. In the harness case, the agent is helping you analyze, solve a problem, and work you through that.
With money, there's some stuff you can leave the agent to figure out, like asking it to do three scenarios when there are a few possible values.
endorsed
Failure modes when you lack validation: checking everything, checking nothing, or having no sense of what 'about right' looks like.
described
Building too many harnesses before using them leads to falling victim to 'axe sharpening' instead of actually using the axe; building a harness from the ground up for every small new thing gets tiring because you spend more time creating tools than using them.
discouraged pattern
Long chat sessions (well over 10 turns) commonly lead to a 'start from the beginning, some stuff we discussed is gone' feeling — a consequence of today's harnesses being designed for linear discussion with everything in one big conversation. Their harness instead breaks the discussion down so users can discuss individual elements without complicating the whole.
If you don't know when to do things in code versus when to build something on the side, you will have a big hit-or-miss ratio.
If there's no feedback loop in an LLM-as-judge setup, the financial decision (or any decision) doesn't improve at all.
described
For forecasting you can try doing the forecast yourself by hand or with a spreadsheet of fancy calculations and likely get a decent forecast; you can get a library from a trusted repo with everything baked in; or you can write code in a more modular fashion with end-to-end testing and good software engineering practices to ensure a good result.
The demo's branching aside feature is designed to show how it could feel if something is designed with the intent that it really needs to be an aside. The speaker has so many 'stupid questions' when chatting with AI that they end up opening a separate chat because they don't trust it won't contaminate what they're doing. Telling the AI 'hey, this is a stupid question. Please don't use it for anything else' doesn't work quite well.
Finance is a super interesting domain. Back in the day, finance felt like just working with spreadsheets — boring work, adding numbers here and there, formulas, whatever. If you ask any accountant, finance is all about creativity and interpretation because they will come up with the weirdest things. In that level of ambiguity, AI is really, really good. You can have spreadsheet bits where the cash flow doesn't quite make sense because some of the revenue was attributed a month later — 'ah, creativity.'
In a branching harness, the model will either assume some decisions and try to verify with you ('is this the number you want to use here?') or try to figure out how things fit into the overall system.
described
When a decision is accepted and marked as made, any subsequent use of that information in the final decision carries the accountability with it.
endorsed
Depending on your mindset, you could look at a finance problem as something very simple an agent can fix within 10 minutes, or as it actually is: a mixture of things that need to run reliably 100% of the time and a lot of interpretation and freedom of creativity in between the lines.
Anekha describes ingesting a large volume of information in her rates career that LLMs could now process better, noting they didn't have the tooling back when she did this 10 years ago.
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.
