Product · OliverDB Swarm

One conductor. A thousand angles.

Swarm turns a question into an investigation. Your agent — the conductor — poses a task; a swarm of small, fast models fans out over your data, each querying it directly and reporting back a grounded, cited finding. The conductor judges. No hallucinated answers — just what the data shows, from many angles at once.

grounded findings adversarial angles policy-scoped cheap fan-out
confirmbee · 07✓ cited

p99 latency rose 41% at 09:12 — the deploy window overlaps.

refutebee · 03✓ cited

No regression in EU — the rise is confined to one region.

alternativebee · 11✓ cited

One tenant’s retry storm explains most of the spike.

→ the conductor weighs all three.
bees query the data themselves — findings come back cited, not imagined
What it gives you

Investigation, not guesswork.

ground

Grounded, not guessed

Every bee queries the data before it answers and reports only what the numbers show — citing them. Inconclusive is a valid answer; invention isn’t.

angles

Adversarial by design

Send angles, not just questions: one bee confirms, one refutes, one hunts the alternative explanation. You get a many-perspective read, not a single model’s take.

economics

Heavy thinking stays with your agent

The conductor poses the task and judges the findings. The fan-out runs on small, fast models — broad investigation at a fraction of frontier-model cost.

govern

Scoped to the lane

Each bee is confined by a policy: it physically can’t read data outside its scope. Point it with a focus; fence it with a key.

connect

Plugs in anywhere

A standard MCP tool for agents, an HTTP service with async jobs and a searchable feed of past findings, or a CLI.

scale

Scale by repointing one URL

The same swarm runs on a laptop or a many-worker GPU box. Dev and prod behave identically — repoint one URL to scale.

Weak bees, made useful by grounding — and a strong conductor.

Part of the Oliver platform

The swarm needs somewhere to land.

Give your agent a swarm.