A Pack

Most "AI research" tools are one generalist agent doing everything badly rather than one thing well: reading, planning and writing are different jobs, and A Pack gives each to a named agent instead of one model context-switching between them. Alpha reads the request and builds the plan; Beta breaks that plan into steps before anyone goes looking for anything. Nothing downstream starts until the pack agrees on what it’s actually hunting for.

The hunt itself is where the parallelism pays off. Scouts range ahead for ground truth, several at once, each chasing a narrow sub-question rather than one agent working the whole brief in sequence; Tracker reads what the Scouts bring back and gives the raw findings shape. Rust/Axum owns the coordination this depends on — dispatch, fan-out, knowing which Scout is still running and which has reported — while Redis Streams is the bus between them, so Tracker starts working a finding the moment it lands rather than waiting on the slowest Scout to finish.

The part that actually keeps the output honest is the layer after the hunt, not the hunt itself. Sentinel challenges any claim that can’t be traced back to a specific Scout finding, and Howler is only allowed to draft the final deliverable from what survives that challenge — so an unverified claim has to clear a gate before it can reach the answer, rather than being trusted because an agent said it confidently. Elder sits underneath as the pack’s memory, carrying lessons from past hunts into the next one. Built in a hackathon window against the Qwen Cloud clock, which is exactly why the verification gate stayed strict: a pack that fails loudly beats one that quietly guesses.