I've made a ๐๐๐๐ special report on how to graft cultural norms into your UX and UI design process.
Upgrade for access to the version with the step-by-step implementation guide from the 'Cultural Considerations in UI Design' link down below.
Reviews and feedback are welcomeI've made a ๐๐๐๐ special report on how to graft cultural norms into your UX and UI design process.
Upgrade for access to the version with the step-by-step implementation guide from the 'Cultural Considerations in UI Design' link down below.
Reviews and feedback are welcome.
@rajesh_praharaj Interested in how you will score the multi-agent versus single-agent comparison. Will you hold the task set and token budget fixed across both setups?
@deviorobert The inspection point is the part I keep coming back to with agent work. What do you record at each step so you can reconstruct what actually happened after a failure?
@GaneshKompella To answer the closing question: our checks only earn trust once they are versioned and run against a fixed set of known failures. Do you have clients keeping a regression set of past agent mistakes?
@Hislordship_9@GoogleAIStudio Good flag. Do you gate model swaps on a fixed regression set of tool-calling cases, or mostly watch finish reasons in production after the switch?
@htahir111 Calibrating the judge against a small human-labelled set is the step most teams skip. How often do you re-check agreement once the agent or the judge prompt changes?
@allysoncodes A benchmark built from your own shipped code also gives you a regression suite for free when you swap models. Do you keep the tasks fixed over time, or refresh them as the codebase changes?
@mervenoyann Using a VLM as the judge inside the labelling to training loop is the part I find most interesting. How do you check the judge's agreement against human labels, so it does not drift towards its own mistakes?
@timmsbuilds Keeping the old links intact while the content was regenerated is the part I would have underestimated. How much human review sits between Claude and what gets published each day?
@ArjitKiled66321 Point 8 resonates. Logging the full input is what makes a bad output debuggable. Do you redact anything before storing prompts, or keep them raw for replay?
@ashanviii Error analysis before LLM-as-a-judge is the order I wish I had followed. Did the book suggest how many manually reviewed traces you need before trusting a judge?
@richieganney Looking forward to the coding agents write-up. I would be curious how you handle cost per task, since that tends to creep up quietly once agents run unattended.
@arielco30 Interesting that you frame it as a shared source of truth problem. Do you version the GTM Brain, so you can trace which change made an agent start behaving differently?
๐จ BREAKING: Google just launched the most powerful coding agent weโve ever seen.
Itโs called Jules.
It reads your codebase, makes a plan, builds features, writes tests and pushes the PR.
No need to co-pilot. Jules just ships.
Hereโs how it works ๐