417 results · Most relevant
In a team, the burden of something being good or correct with vibe coding ends up in the hands of the reviewer — demoralizing, since they're handed stuff that looks about right but it's their responsibility to ensure it is. Team members also pass blame for bad decisions on the basis of 'no but Francesco will review it.'
The presenter recounts that a couple of times he thought he had given all the necessary information, the agent returned a formula (even nicely formatted), but he knew the result was wrong and started from scratch. A sign something is off is when results look reasonable but running the same thing again gives a different result — then you need to do it differently or better. An audience member gave an example of a number (92) being read inconsistently (92 then 99) and said they just don't trust chat-based calculations.
When reviewing 100 tabs of agent changes, most feel like 'come on, this is a one-line change' but some feel like 'this is important, thank God I was looking at this' — which is why review gates matter.
People pass the blame for bad decisions onto whoever was supposed to review, on the basis of 'no but Frenchex will review it' — and it happens. With vibe coding, the burden of something being good or correct ends up in the hands of the reviewer, which in a team feels demoralizing because you're given a bunch of stuff that looks about right but it's ultimately your responsibility to make sure it is right.
Francesco Galletta lists testing — a decades-old practice — planning with your agent, doing evals, traces and observability, and loops and graphs as components of building trust in agent systems, noting he could keep naming components.
Gaining trust is less about the AI implementation itself and more about aligning on what outcome you're trying to drive with the AI.
The harness is an example of how to apply the principles, but everyone should be thinking about their own use cases, their companies, and the agency they are building or using, and how to connect the dots.
Felippe warns that bad decisions carry forward into your other decisions, so you need to know that.
A common data-engineering failure is being never sure whether a number is correct — 'it sounds like it is, but we're never sure.'
To set up the workshop materials, scan the QR code or go to the repo and follow the setup instructions — you don't need to execute anything just yet, just get the repo into your preferred setup. The repo contains a guide for the workshop, the wedding example, the harness itself, and a folder for anything attendees build that drifts from the main flow. If the setup works, you should see the UI open and a success message in the terminal; the second command opens a UI designed to show how the harness behaves and should feel in the real world. To switch between the wedding example and your own use case, delete the database in the my/var folder, then reopen the UI with the previous command.
Anekha describes building a vision language model to understand and process messy PDF data from clients, and taking client data uploaded to their cloud version of Moshi, breaking it down, and building it up in their database.
When distributing agents to internal users, a user's user-level claude.md file may contain instructions that go against what you're telling the agent to do.
described
At Monzo, there is an internal contribution mechanism for agent extensions where multiple teams can share what works internally, like plugins and skills.
described
Ask the finance team whether they supply golden answers or a little bit of a harness, so you know you're doing things correctly in terms of things to balance with. Think about the trajectories of a user making a very final decision and look at the trajectories — what tools and information were used.
The hosts explain that if you change multiplication to an aggregation in the spreadsheet you may make the quick edit, save the file, and see success — but you will sabotage the analysis, because the harness's tests will stop you from reaching any conclusion of the analysis you wanted to do.
Maintain a semi-strict monthly close of budget and categorize everything in your statements.
As a BI analyst, Francesco's query was good, the number seemed reasonable, he checked everything and nothing looked off — but when he presented the dashboard to a stakeholder, the stakeholder immediately said 'that's not correct.' The stakeholder knew instinctively that the number can't be right because it meant the information in another place was incorrect — they just had an idea of how many cells were made in that year.
The speaker's fiancé invited three more wedding guests without telling him, so his whole model of the number of people invited was wrong. Having a wrong assumption somewhere and not knowing where it came from can be fatal for any financial analysis.
When building the parts that make up the behavior of your harness, avoid oversharing because of IP.
You can easily fall victim to doing a lot of axe sharpening and not actually using the axe.
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.
