@finalfantasyma2 building a marketplace myself and this is the bit the big platforms keep getting wrong. what would a platform need to promise for you to trust it?
@amelia_tweetz the more open a marketplace gets, the less you trust any single listing. smaller communities where you know who's selling fix a lot of this
@systemdesignone how did the 4 agents share context? we tried splitting the build across agents and went back to one loop because the helpers never had the full picture.
@yituihi we did the basics for this. every long job has a tracker file in git, and each new session starts by reading it and picks up at the first row that isn't done. no memory layer
never done a proper intro on here, so here it is.
we build resale marketplaces. around 50 fashion brands run their own second-hand store on our platform, and we're now building Clubs which is an app where fans can buy, sell and chat with each other.
most of my job now is building the system the agents work in.
will be posting about:
- how we run coding agents in production. what works, what breaks, what it costs
- marketplace data. supply, demand and what makes people come back
- where the two meet
If you're building with agents or running a marketplace, or none of those, say hi.
hiya mate. there's a few things we've done that have really helped:
- built a nodejs service in the cloud that uses personal access tokens with guardrails. much more bang for your buck
- fallback to api token usage with cost caps e.g. "max $20"
- signed up for claude for startups. $500 in token usage per month
- build a model harness with Jev. Jev basically decides what model to use for each task. hugely reduced our token usage as we're using lighter models for lighter work now.
happy to share more on any of these!
@Pavan_Belagatti yep. we've flipped ours over the last few weeks. one agent in one loop, no swarm, and we sit at the two ends. we approve the test scenarios at the start and read the evidence at the end. basically TDD on loop.
@kmeanskaran great list. I'd push back on one tho - we do let agents run commands and block the few that can hurt with deterministic hooks (bash scripts).
@hwchase17 agree the context is the hard part. ours is a git repo of markdown that every agent run pulls first. conventions, postmortems, and a ledger of mistakes our agents made.
the hard part is now managing the bloat ๐คฃ
@Miaarteh@coderabbitai we've switched to a more TDD approach, and it's been major for us. it's made our code way more deterministic and has massively reduced reruns.