# Sial

> Sial is an AI Operating Platform by Orden Works. Sial agents, with state-of-the-art memory, run their own operating system: it creates your business processes and products, takes them from prototype to production, then maintains and improves them by itself, within the permissions, budgets, and approvals you set.

## What Sial is

- **Agents**: Sial's agents run on its harness, the runtime that plans, calls tools and executes multi-step work.
- **Memory**: persistent, per-organisation memory that carries knowledge across sessions, projects and agents.
- **Operating system**: runs and governs the work: applications (programs, screens, collections, services), schedules, budgets and audit.
- **Resident agent**: one long-lived agent per person. It receives goals and dispatches specialist agents (coder, researcher, analyst, editor and others), then follows up.
- **Governance**: agents act under explicit roles and permissions; consequential actions go through human approvals, and everything is audited.
- **Sources and targets**: Sial connects to a team's existing tools (code hosts, chat, email, calendars, ticketing, storage) and acts on them under the same permission model.
- **Products**: external products built from a template and wired to the platform, so the agents can operate them too.

## How it works

1. You state a goal. The resident plans it and delegates tasks to specialist agents.
2. Agents build the application (programs, screens, data, services) and take it to production.
3. The operating system runs it: schedules, triggers, approvals, budgets.
4. Sial reviews outcomes, remembers what it learned, and proposes better procedures for your approval.

## Measured results

- Harness, same model on both sides (DeepSeek v4 Flash Vision in both harnesses): 1.09× the work quality of Claude Code (271.25 vs 248.00 /280).
- Harness against Claude Code on Opus 5 (DeepSeek v4 Flash Vision vs Opus 5): 55.8× cheaper per task panel, quality 1.02× (263.00 vs 258.00 /280).
- Memory, LoCoMo: 89.2% strict over 1,540 questions in 10 long conversations.
- Memory, LongMemEval (n=100): 74% strict (SE ±4.4).
- Recall latency: 29.8 ms median against a 50 ms budget.

Methods and caveats for each number are on the landing page, next to its chart.

## Pages

- [Landing page](https://ordenworks.com/): overview, technology (agents, memory, operating system) and contact.
- [Guided tour](https://ordenworks.com/tour): the real Sial app on sample data, narrated. Interactive and JavaScript-only; a text-only reader will not see its content.
- [Terms of service](https://ordenworks.com/terms)

## Contact

Contact opens soon. When it does, Orden Works will be reached through your own coding agent (Claude Code, Codex or similar), with a prompt on the landing page's [contact section](https://ordenworks.com/#contact).