We had to choose: build @trysecretsauce with templates or go fully agentic. Templates would have shipped faster. Here's why we didn't. Templates put a ceiling on what AI can do for your brand. They're fast to produce and consistent in output - but consistency comes from constraint. You can only get what the template was designed to give you. And the template was designed for everyone. If your brand doesn't fit, you adapt the brand. That's backwards. An agentic system has no ceiling. You get the best practices baked in, but you're not borrowing someone else's aesthetic. The output is shaped by your brand - your colours, your tone, your vibe - not by what the template allows. If you can describe what you want, you can make it. The tradeoff is that it's harder to build, harder to control, and harder to make consistent at scale. Those are real problems. They're also the problems worth solving. We didn't want to build the ceiling.

We first built SecretSauce Labs as a pure chat product. It was the wrong call, but I'd do it again. At first, people came to the platform and prompted to create content. The problem showed up fast: people didn’t know what to do with an open canvas. But I’m glad we made that mistake because open chats reveal what people actually want. If we added buttons and created a dedicated UI too early, we wouldn’t have such strong insights from our users. So start conversational while you’re still learning about your customers. Then build dedicated UI once the patterns become clear. This is how we’re approaching every new feature going forward. Your product will become easier to use because it’s actively learning which questions users keep repeating.

A brain simulation living in a world simulation.

If you’re building generative media or marketing agents in-house, here are 5 new releases I’d look at this week: 𝟏. 𝐎𝐩𝐞𝐧𝐀𝐈 𝐃𝐞𝐜𝐢𝐬𝐢𝐨𝐧𝐬 𝐀𝐏𝐈 Picks the right model, tool, or action in near real-time, up to 10x faster than running the same decision through GPT-6 Luna. Worth exploring for choosing tools or checking creative outputs against specific requirements, without asking a general agent to reason through everything. xitter.cf/OpenAIDevs/status/2107… 𝟐. 𝐁𝐮𝐢𝐥𝐝𝐢𝐧𝐠 𝐚 𝐌𝐞𝐦𝐨𝐫𝐲 𝐏𝐢𝐩𝐞𝐥𝐢𝐧𝐞, 𝐰𝐢𝐭𝐡 𝐃𝐞𝐱: What should your agent remember from the work? I’d start with the corrections your team keeps repeating. Just make sure a temporary campaign choice doesn’t become a permanent brand rule. xitter.cf/i/broadcasts/1nxnRBQgm… 𝟑. 𝐅𝐋𝐔𝐗 𝟑 𝐯𝐬 𝐍𝐚𝐧𝐨 𝐁𝐚𝐧𝐚𝐧𝐚 𝟐.𝟏 Ori Silver’s comparison flags degradation across repeated edits. Test the whole revision process on your own assets, not just the first generation. Does the product still look right after everyone’s feedback? xitter.cf/OriSilver/status/21075… 𝟒. 𝐍𝐚𝐧𝐨 𝐁𝐚𝐧𝐚𝐧𝐚 𝟐.𝟏 Promises better instruction following, text rendering, and lower pricing. Try rerunning your difficult briefs with it. Measure the cost per approved asset, including retries and cleanup. xitter.cf/_philschmid/status/210… 𝟓. 𝐌𝐞𝐫𝐠𝐞 𝐆𝐚𝐭𝐞𝐰𝐚𝐲 𝐄𝐯𝐚𝐥𝐬 Test a new model on your agent’s actual tasks before switching to it. Load the test set with the awkward briefs and the failures you already shipped, not the examples that pass. xitter.cf/merge_api/status/21074… And if you’re the person trying to make this work inside your company, I’d be curious where you’re getting stuck. This is the kind of work we focus on with SecretSauce 💗

To be frank, this is the UX that Codex should have had on ChatGPT, not this crazy, confused experience that we went through.

we're all just figuring out the same things, really

State of AI today: currently fixing non-deterministic models with fast probabilistic ones.

That is just SO good.

codex is down

You can't have a decent conversation with it AND it will 100% whoop your ass.

TLDR: Instinct persistent memory is markdown files, git-tracked and interlinked, with the LLM maintaining the knowledge + cron job to maintain ~once a day. KW matching/grep for retrieval, no vector search or BM25. Simple and elegant. Very cool! Thanks @DhravyaShah !

Some will claim this is AI.

As a creative, I am incredibly relieved to hear that Joel Kuwahara did not, in fact, lose all artistic judgment, moral integrity, hard-earned skills, standards and his will to live when he started using a new piece of technology. Incredible news for all us creatives using AI, who were in great danger of absolute moral bankruptcy.