Skip to content
aykol.fyi
← All posts

BrugBot · RedHorse3 min read

How we run 12 agents on free model tiers

A twelve-agent debate sounds like dozens of model requests per analysis. In BrugBot an analysis averages about five. How early exits, collectors that spend no requests and a provider chain got us there.

The criticism we hear most about multi-agent systems is cost. If twelve agents debate a token and each calls a model once or twice, twenty or thirty requests per analysis add up fast. Our constraint was tighter still: we wanted to run both projects on free model tiers.

Today BrugBot averages about five model requests per analysis. Three things made that possible.

1. Early exit

The full desk flow is: four analysts, a bull and bear debate, a judge, then a trader, a risk team and a portfolio manager. But if the judge says "hold" or "sell", it's already clear there will be no buy. In that case the trader, the three-person risk team and the portfolio manager never run.

Most analyses end there, because most scanned tokens genuinely aren't worth buying. Early exit saves both minutes and requests; the full flow only runs for real candidates.

2. Work that doesn't need a model doesn't go to one

Something that can be done with AI isn't the same as something that needs AI. These run without a model:

  • The collector records trending lists, new tokens, signals and smart-wallet buys. It's the raw material for backtests, and it spends zero model requests.
  • The risk gate checks liquidity, wallet distribution and authorities with rules. A token that fails the gate never reaches the debate; hundreds of candidates a day stop here.
  • The rules for size, budget and exit plans live in code. Models are asked for judgement, not arithmetic.

The same principle holds in RedHorse: match probabilities come from a statistical model, and AI only reads the news, runs a second check and writes the advisor's answers.

3. A provider chain, and working without AI

Both bots share the same AI connection: NVIDIA's free endpoints first, then OpenRouter's free models. Every service has a daily request quota, and the quota is spread across the day, so no service can burn its allowance in the morning.

More importantly, both keep working without AI. If no model answers, BrugBot's approver stays on the safe side and rejects, and RedHorse's advisor falls back to a short rule-based answer. Free tiers slow down or drop out from time to time; the system had to treat that as normal, not as a failure.

Short context, good context

Sending each agent a long history on every call makes requests expensive and often makes answers worse. Instead, each agent sees:

  • Its own duty card: its goal, its inputs, who it reports to and what it must not do.
  • The team's chain: who runs before it and who runs after.
  • A few current lessons that the student container distils from settled decisions.

That gives every agent a short context that is actually useful.

What we measure

Every day we log how many requests were spent, the average per analysis, and whether analysing the same token twice produces the same decision. The biggest risk with free models is inconsistency, giving different answers to the same question on different days. So far repeated analyses agree, but that measurement stays on.

You can watch what the agents are doing in this flow on the map in the BrugBot demo.