The One-Minute Tour

Sixty seconds. No jargon.

Overview

Remember three ideas: Aegis checks actions rather than trusting generated text; a human approval is bound to one exact action; and every protected decision leaves independently verifiable evidence.

Status: These known-agent integrity capabilities are implemented. Cage, sensor, forced egress, broker enforcement, and multi-replica HA remain Partial.

The problem

Companies are handing AI agents real power: credentials, APIs, code, money, customer data.

The risk

Agents do what text tells them. Attackers write text. One poisoned ticket, issue, or email can turn your agent into their agent — and afterwards, nobody can prove what actually happened.

The solution

AegisAgent lets you run AI agents without losing control.

Before an agent does something risky, Aegis:

1. Checks it — deterministic policy, based on where the request came from.
2. Controls it — risky actions wait for a human; the approval locks to the exact action.
3. Records it — every decision becomes a tamper-evident receipt.
4. Proves it — the receipt chain shows exactly what happened, to anyone.

And if an agent misbehaves, the built-in SOC can freeze, quarantine, or ban it.

The simple picture

flowchart LR
    A[AI agent] --> C[Aegis checks]
    C -->|safe| RUN[Action runs]
    C -->|risky| H[Human approves exact action] --> RUN
    C -->|dangerous| STOP[Blocked]
    RUN --> P[Provable receipt]
    P --> W[SOC watches and can stop the agent]

Next

• What it is, in full: What_Is_AegisAgent.md
• Why it works this way: Why_AegisAgent.md
• The mechanism, step by step: How_It_Works.md
• I'm a developer — integrate now: onboarding/For_SDK_Developer.md
• The full story with code references: Last_Mile_System_Walkthrough.md

Try It

make demo

The expected outcome is a blocked malicious action, a rejected action swap/replay, and a verified receipt chain—not merely a running dashboard.