One fish.
You, one chat window, one job. You hold the context, re-explain what you meant, and paste things between tabs. It works, until the job gets bigger than the window.
We're developers. We were working on a hard problem: a new kind of AI, not another language model, and the work needed agents that could run for days, check each other's work, and not bankrupt us doing it. Nothing out there did that. Chatbots forgot. Swarm frameworks died halfway or burned the budget. So we built the tool we wanted: complex, long-running, quality-driven swarms with full access to our machines, no life span, and a meter on every dollar.
It was supposed to stay ours.
Then someone asked what we were using. They wanted it too.
Then someone else.
And someone else.
We would have kept it a secret. Our bosses made us share it. Hence GoatFish.
In 2018 a group of scientists watched Parupeneus cyclostomus hunt and wrote the moves down as an optimization algorithm. One fish chases. The rest block the exits. If a blocker finds the prey first, it takes the lead. If the reef is empty, everyone moves. It turns out to be a good way to run a team of AI agents, too.
You, one chat window, one job. You hold the context, re-explain what you meant, and paste things between tabs. It works, until the job gets bigger than the window.
A real job is never one thing. The report needs its numbers checked, the numbers need their source, the source needs cleaning up. One assistant chasing all of it at once catches none of it.
The team splits into small squads. In each one, the agent with the best read on the problem leads. The paper calls it the chaser. It makes the long, confident moves.
Blockers don't chase. They stand where mistakes would slip out: checking a figure against its source, reviewing the draft, keeping two agents from editing the same file. In GoatFish, those are the checker and reviewer agents working beside the lead.
When a spot is picked clean, the whole team moves to wherever the best fish is. In GoatFish, when an agent runs out of memory or budget, the job hands itself to a fresh one, notes included, and keeps going.
If a blocker finds the answer before the leader does, it becomes the leader. In GoatFish that's a second model reviewing the first one's work, and sometimes being right.
chaser x′ = x + α · ( u / |v|^(1/β) ) · ( x − x★ ) # Lévy flight blocker x′ = D · e^(bρ) · cos(2πρ) + x_chaser # log spiral
Zaldívar, Cuevas et al., A novel bio-inspired optimization model based on Yellow Saddle Goatfish behavior, 2018. The fish were not credited.