Aegis AI is a controlled, autonomous multi-agent intelligence platform for defence and other high-assurance environments. It is designed to plan, act and improve on its own — with every action passing through a military-grade safety architecture you stay in command of, on your own hardware, with no cloud dependency. Our most ambitious system, in active development.
Autonomy is only useful if you can trust it. Aegis is built to govern every action through a layered safety stack — the difference between an AI that acts, and an AI you can hand authority to.
An atomic halt that stops the system instantly, with revival gated behind operator authority. Role-based control — system-admin, operator, observer, autonomous — decides who can do what.
A formal escalation state machine (Normal → International) with a risk multiplier. Higher-stakes actions demand higher assurance before they ever execute — no silent escalation of capability.
An immutable identity core with a fixed set of values that every action is checked against, in two layers. The system cannot quietly redefine its own purpose or step outside its mandate.
A multi-component hallucination-detection pipeline, prompt-injection blocking and PII redaction — so outputs are filtered and adversarial input is contained before it reaches an agent.
A single orchestrator routes each task across a swarm of specialist agents — security, reasoning, engineering and analysis — backed by a six-layer memory.
One entry point routes each task to the best-fit agent(s) — a SecurityGuardian for threat detection and policy enforcement, plus reasoning, engineering and analysis specialists — under one coordinator.
It runs on its own initiative: a perceive–plan–act–reflect–learn cognitive cycle that senses a situation, plans a response, acts, then reflects and learns from the outcome — without waiting to be prompted.
A hybrid memory spanning a structured store, vector recall, an append-only ledger and episodic layers — so the system remembers context, decisions and lessons across restarts.
Designed to run on open models on your own hardware — edge devices to a private cluster — with no cloud dependency, for environments where data physically cannot leave the building.
A continuous health-and-drift watch: the engine spots configuration and behavioural drift, proposes remediations and applies them under approval — with a circuit breaker that isolates faults before they spread.
A self-model tracks its own capabilities, limitations and live state — so it attempts only what it can actually do and never claims a skill it lacks. A world-model holds beliefs with confidence and provenance, updating them as evidence arrives rather than pattern-matching text.
A unified awareness layer lets the agents reason as a single mind: when one learns something, all of them know it; when one fails, all of them adapt. The swarm thinks as one.
One codebase on every node — olympus, talos, an edge Pi. Each profiles its own hardware and the strongest self-elects as the brain; there is no central master. Lose a node and the fleet re-elects and carries on, and a kill switch on one cascades to all.