The research engine
for frontier R&D.

Five roles. One coherent investigation.

Most research tools give you a search result or a summary. SwarmOS runs adversarial structure and delivers a maintained investigation that updates as evidence evolves.

One question. A living investigation.

Point SwarmOS at a hard technical question. The swarm decomposes, gathers evidence, disputes weak claims, and synthesizes — then keeps the answer current as the landscape shifts.

 1from swarmos import Investigation
 2
 3# A question that breaks one-shot LLM workflows
 4inv = Investigation.create(
 5    objective=(
 6        "Does Flash Attention 3 hold its claimed 1.5x–2x speedup "
 7        "in production workloads beyond A100/H100 benchmarks?"
 8    ),
 9    domain="ml_infrastructure",
10    depth="comprehensive",
11)
12
13# planner → scouts → extractors → skeptic → synthesizer
14for event in inv.stream():
15    if event.type == "claim":
16        print(f"[{event.confidence:.2f}] {event.text}")
17    elif event.type == "contradiction":
18        print(f"⚠  conflict: {event.summary}")
19    elif event.type == "answer_updated":
20        print(f"\n→ Answer v{event.version} (conf: {event.confidence:.2f})")
21        print(f"   {event.text}")

Start your first
investigation.

Your hardest open research question — answered with adversarial structure, versioned answers, and a Skeptic that doesn't let weak evidence slide. Frontier R&D deserves better than a chat transcript.