Benchmarks / Fleet residency
What a dormant agent costs
A million dormant agents written to disk and read back: 477 bytes of state each, 455 MB for the million, and a resume in 0.034 ms. Measured on a laptop, against a competitor figure that was modelled and never billed.
- harness
- hanzoai/cloud · bench/fleet
- our paper
- What a Dormant Agent Costs
node fleet/fleet.mjs /tmp/fleetprints the table on this page — transcribed from bench/README.md, the pass of 2026-09-07
What is measured
A million agents are created, written to disk, and read back. What is counted is the state a dormant agent occupies — the row that has to exist for an agent to be resumable at all — the rate at which those rows can be written, and the time to bring one back. Nothing is running while the measurement is taken, which is the point: an agent platform’s floor is what it costs to keep an agent that is doing nothing.
Results
one million agents · M-series laptop · the pass of 2026-09-07
| measurement | Hanzo, measured | the published figure |
|---|---|---|
| state per dormant agent | 477 bytes | ~1 MB, a competitor’s own modelled assumption |
| one million dormant agents on disk | 455 MB | ~977 GB at that assumption |
| dormant agents written per second | 309,789 | — |
| resume one, in process | 0.034 ms | — |
The right-hand column is not a competitor’s measurement. It is the per-agent figure a competing platform assumes in its own published arithmetic, marked there as modelled and never billed. Ours is a directory you can list. The distance between the two is about 2,200× and most of it is the difference between an assumption and a file.
What this does not measure
A dormant agent is not a running one. This number is the floor, and the floor is the easy part. What a live agent costs is a different measurement with a different command — 601 bytes of heap and 187 ns to spawn — and what it costs to give one a shell is a third one, at 150.8 ms for a cold container.
The line that grows is not agents. A dormant agent stays 477 bytes; its transcript does not. At 2 KB retained per call, a billion calls a month is 22 TB after a year, and history eventually exceeds compute. Any cost claim built on the row size above is a claim about the wrong axis, and this page would rather say that than let the 477 be read as the whole bill.
One machine, one day. This is a single pass on one laptop. It is reproducible by one command and it is not a fleet of machines under load, so read it as a floor that has been established rather than as a service level.
The other benchmarks
LoCoMo · MemoryAgentBench · LongMemEval · RepoBench-R · LoCoMo · subject scope · LoCoMo-Conv · Live agent footprint · Sandbox cold start · Inference vs llama.cpp · GPQA-Diamond · all of them, and the head-to-head
harness and the committed table: hanzoai/cloud · bench/fleet · the paper: What a Dormant Agent Costs