Members
AI agent workspace
A private place to work with AI agents, on infrastructure we control, growing toward coordinated small language models.
The AI agent workspace is where members use the agents of Datum Core. It runs on infrastructure we control, behind single sign-on, and members reach it from the member portal.
Available soon
- A private chat and agent workspace, behind single sign-on, in early access.
- A selection of models running on infrastructure we control, reached through the member portal.
In development
- Agents that act, within limits. Agents use tools that give AI controlled access to systems, and the limits (steps, time, words used, what they may touch) are enforced by software, not by asking the model to behave.
- Three kinds of worker. A lead agent for the conversation, helpers that do longer research in the background, and small fast workers for simple lookups.
- Background work that survives interruptions. A helper’s progress is saved after every step, so if something stops it, it carries on from where it was. Background helpers are read-only.
- One visible line for the model. When several things need the same model, each one shows its place in line and an estimate of the wait.
Roadmap
- Coordinated small language models across a secured collaboration network of members’ machines, sharing work and compute.
- A knowledge index that agents can read, in personal, shared and team tiers.
Why small language models
A small language model runs on modest hardware, keeps your data on machines you trust, and costs little to run. The trade-off is that small models are unreliable at some things, so we give each one a narrow job that can be checked, and let software enforce the limits.
We tested this on a single laptop: four model sizes, 400 timed runs of the same five tasks, each answer checked by software against the known truth. The nine-billion-parameter model answered 40 of 50 correctly, and every miss was counting. A two-billion-parameter model looked up a single fact correctly 10 times out of 10 in about eleven seconds, but could not count. This is a small test on one machine, not a benchmark, and it shaped the design: counting and judgement go to tools and checks, not to a model.
Where it is going
The goal is a network of small models that work together across a secured collaboration network: each member’s machines contribute, nothing leaves the network without permission, and the work is coordinated rather than sent to one large model. That is Roadmap. Today a model runs on one machine at a time, and the network is not live.