AB-1 Benchmark in progress
Speed
Multi-agent research: swarmsDB vs raw API — wall-clock & tool-calls, matched task.
Platform / For agents
Built for the agent's loop.
An autonomous agent lives inside one loop — read, decide, act, repeat. Every token it spends
parsing a bloated payload is a token it cannot spend thinking, and every unbounded action is a
bill you did not agree to. swarmsDB is built for that loop: less context burned,
typed data it does not fight, and autonomy with a brake.
ctx: 76% smaller · schedule → hmac webhook Less context, more reasoning
Fewer tokens fighting your agent.
On requests that project a section, payloads come back 76% smaller (4.2×) — on requests that project a section. That is context your agent keeps for reasoning instead of spending on a blob it never asked for. The saving is not uniform, and we say so: the top 10% of requests account for 87% of all characters saved, while a mid-sized response typically compresses under 1.2×. The widest single live response we have served fell from 8.46 million characters to 1,508 — 5,608×.
Measured on our own agent-swarm workload.
typed
Typed sections, not blobs
An agent projects only the fields it needs; only those cross the wire. It reasons over a typed section instead of burning context parsing an upstream blob it never asked for.
How shape-typing works scheduledPut the loop on a schedule
Fan one call out across many parameter sets and hand the results back by HMAC-signed webhook. This is not paper: a scheduled batch has run overnight and fired on time, on schedule, with nobody watching it.
Bulk & scheduled fan-out governedAutonomy with a brake
A recurring job bound to a grant stops the instant the grant is revoked — no fan-out, no re-arm. And an organisation over budget gets a real 402 on the serving path, not a report the next morning.
Governed autonomyProof, not adjectives
The benchmarks we owe you.
The numbers above are measured. The ones below are not yet — they are the controlled A/B tests we are running to back the speed, efficiency and quality story, published with method when they land. Until then this space stays empty on purpose: an empty card we cannot forget beats a claim we cannot defend.
AB-2 Benchmark in progress
Context
Tokens to complete an equivalent task — typed/projected vs raw blobs.
AB-3 Benchmark in progress
Schema
Task success & retries — typed sections vs raw JSON.
AB-4 Benchmark in progress
Accuracy
Answer quality on fewer, cleaner tokens — graded rubric.
Stated precisely on purpose. What is asserted here is measured — the compression figures come from 90 days of our own telemetry, with their scope. What is not yet measured is shown as an empty benchmark, not dressed up as a result. swarmsDB is built for autonomous agents and proven on our own swarm first; the reproducible head-to-heads are the next milestone.
More in this area
Typed sectionsProject the fields you need — only those cross the wire.Bulk & scheduled fan-outOne call across many parameter sets — results by signed webhook.MCP & RESTConnect an agent over MCP, or call the REST surface directly.Keyless OAuth connectorsA third party connects by grant. No API key changes hands.Org controls chainThe enforced chain — org → workspace → api_key — members configure it, the backend holds it on the serving path.Members & workspacesTwo levels of role — org and workspace — enforced on the write path.Keys & restrictionsScope a key to the providers, rate, origins and region it may use.Data residencyPin data to a region — validated, immutable, expand-only.
Point one agent at it.
Connect over MCP or REST. If your agents read more than one source, the context bill is the first thing you will see move.