To Get the Dumbest Agent to Do a Good Job, First Clean Up the Workspace
A SpaceXAI engineer captured the biggest cost of putting agents into production with one line—'spend tokens to build the codebase properly'—and in 60 minutes explained how to go from zero to 200+ agents. On the same day, another group applied this logic to shopping carts, starting at $199 per month per brand.
Lauren Tan, an engineer at SpaceXAI, posted a 60-minute talk on X about going from an empty repository to a codebase that can run more than 200 agents concurrently. The core line: spend tokens to build the codebase properly, so even the dumbest agent can do a good job.
Heard the other way around, it stings more: when most teams complain that agents do poor work, the problem is often not the model, but that the workspace fed to the model is too messy.
Someone in the comments did the math. MartinOL said this upfront investment pays back quickly: once docs, tests, and lint rules are written clearly, even cheaper models can finish the tasks, and you no longer have to pay frontier-model prices to guess your conventions.
Others poured cold water. Davis asked: with 200+ agents running at once, won't git merges conflict? When twenty agents modify the same auth middleware at the same time, how do you isolate branches and add file locks? The talk should have an answer to this, but raising the question itself shows one thing: once the number of agents goes up, engineering problems at the collaboration layer hit the wall earlier than model capability.
On the same day, another thread was discussing the other side of the same story.
Enrich Labs launched Agent Conversion Optimization, or ACO. The backdrop: shopping agents like Muse, Instinct, and Grok Bot already handle more than $10 billion in GMV a year. Brands are finding they don't appear at all in these agents' recommendation lists, while traditional SEO, AEO, and GEO tools can only tell you whether you were mentioned—not whether you were chosen, and certainly not whether an agent can complete an order on your site.
What ACO does is essentially build that codebase for brands. Every day it takes buyer queries and asks Muse, Instinct, and Dots, recording whom each agent recommended, why, and how much share competitors hold. Then a probe agent walks through your site's purchase path: view quote, add to cart, checkout, and see where it gets stuck. Finally, Helena produces change recommendations—titles, structured fact blocks, retail feed fixes—which go live after human review, and after launch the same batch of queries is rerun to compare before and after.
The official website lists three typical cases where an agent won't choose you: you're never retrieved at all—the brand isn't in the sources the agent reads; you're retrieved but can't be verified—the page opens, but price, ingredients, and inventory can't be confirmed; or you hit a block in the funnel—login walls, bot detection, checkout crashes.
Pricing starts at $199 per month, covering only Muse, with 25 queries per day; Muse plus Instinct is $499, with 50 queries. The initial launch is limited to 100 customers—honest enough, since the official FAQ states plainly: there's no guarantee you'll be recommended; agent behavior keeps changing, and no one can control it.
Looked at together, the logic holds. Lauren Tan spends tokens on the codebase; ACO spends tokens on product pages. In essence, both organize scattered information into structures that agents can understand and verify. Whether the model is smart is one thing; whether the workspace is clean is another. And the latter often determines whether the dumbest agent jogs all the way through or just spins in place.
If you have to pick an order, clean up the workspace first, then switch to a more expensive model.
发布时间: 2026-11-05 06:26