For internal teams · business
Provide agent swarms inside your organisation, with the governance already written.
People across the organisation want agents doing real work. Security, risk and legal want to know what those agents can touch. This gives both: a hive with a written mandate per role, enforced, with receipts, and data that lives only as long as the run.
Built on rUv's tools: Ruflo (MIT, by rUv / ruvnet), an agent meta-harness for Claude Code and Codex with a queen-led hive-mind and a policy engine that runs in legacy, observe or enforce mode and writes a hash-chained receipt for every decision. We build on it; we do not sell or own it. Ours: sgit vaults and Agent Behaviour Policies (Apache-2.0).
Who this is for
- Platform and AI enablement teamsasked to offer agents as an internal service.
- Security and risk teamswho need the mandate, the decisions and the data path in writing before they say yes.
The problem, in your words
"Everyone wants agents. I am the one who has to explain what they did, with whose data, if something goes wrong."
Governance
A written mandate per role
An Agent Behaviour Policy for each role: what it can reach, what it is mandated to do, the barrier between the two, and the person who owns it, valid until a date.
Enforced, not advised
The mandate compiles into Ruflo's policy engine in enforce mode. On a fresh project the engine runs in legacy mode; observe sits between the two. In our trace, enforce stopped 6 of 11 actions: 4 denied, 2 sent to a named person. POC1 →
Receipts
Ruflo writes a hash-chained receipt for every decision. They go into the run's vault with the policy that produced them.
No network by default
Runs get no internet. Materials are provided up front, and the policy denies egress.
Data
- Encrypted vaults. Files are encrypted on the client before upload; the server stores ciphertext under opaque ids.
- Ephemeral per run. One vault is created when the run starts and holds everything the run made.
- Deleted, and the delete verified. After the run the vault's objects leave the server, the key is destroyed, and the delete is checked: no stored objects left, and a clone after the delete fails.
Every run so far, with its numbers, is in the ledger.
Deployment
On its own machine
Dedicated compute: one AWS EC2 instance per task, small, medium or GPU, started for the task and stopped after it, with the vault created inside it. Designed, not built yet: SG Compute is not connected to runs.
In your own environment
The maintained repository: the code that runs this, kept current, set up for your stack. Runs on Claude Code, on AWS, or anywhere you run code. Anybody who implements this implements sgit vaults and Agent Behaviour Policies.
Not automated end to end yet. A person on our side starts each run and watches it to the end. The runs on this site used a local vault server on the machine that ran the hive, so their timings are local numbers; the workers were Claude Code subagents, with Ruflo holding the queen, the roles, the tasks and the broadcasts. Dedicated compute and the materials sandbox for customer uploads are not built yet, and SG Compute is not connected to runs: when we reply to an order we say what we can do today.
How we help you start
Expert time, mixed with runs. A couple of hours to review your first policy. A day to run the first hive with your team. Several days to set up your deployment and hand it over.
Offers that fit
Prices are in the shop. Mix and match: put what fits in the cart and send it; there is no payment step, the order comes to us by email and we reply.
Where this sits
Dedicated compute: the parts you run yourself and the parts you buy. Read the map with the business model.
Talk to us
A live agent answers; a person follows up. Every contact and every order from the shop goes to this address.