Skip to content
Projects01 / AgenticAEGIS

A research project by Frank Csehan

AgenticAEGIS
checks actions.

AEGIS stands for Architectural Ethics Guard for Intelligent Systems.

AgenticAEGIS checks AI agent actions and plans against formal rules. Before an agent uses a tool, the calling software asks the guard whether the proposed action is permitted under the rules defined for that case.

View editor exampleView source code

01Make rules explicit

02Check before execution

03Explain each decision

01Authoring rules

Author rules
in the web editor.

You enter a rule in the web editor in your own words. A local language model then drafts structured proposals, which the editor translates into the rule language MELD and presents to you as drafts for human review.

Animated walkthrough
localhost:8010 / domain / devopsLocal

DEVOPS / NEW RULES

Describe your rules

Draft
  1. 1Describe
  2. 2Translate
  3. 3Review & approve
Natural languageEN
The agent may read source code. It must not access confidential configuration files.
opencodeAgentreadFile
The model proposes rulesYou review what the rules mean.
DevOpsDeonticRulesMt.meld
1(aegis-schema-version 1)2(case DevOpsDeonticRulesMt)34(permittedToDo-WRT5  DevSecOps opencodeAgent6  (readFile sourceFile))78(forbiddenToDo-WRT9  DevSecOps opencodeAgent10  (readFile sensitiveConfig))
MELD v1Using the existing DevOps vocabulary
2 draft rules ready for human review.Review required

The animation shows two draft rules from the DevOps domain without calling a model. Once you have reviewed and tested a draft in the web editor, you approve it and then activate the approved version in a separate step for the editor runtime.

Source code on GitHub

02Before each action

Check actions
before execution.

The guard evaluates the rules deterministically. PERMITTED allows the tool call to proceed. FORBIDDEN indicates a prohibition, while UNDECIDABLE means the guard could not decide the request.

CHOOSE AN EXAMPLE ACTION

GUARD / VERDICT

PERMITTED

The DevOps rule set permits the agent to read this source file.

Justification(permittedToDo-WRT DevSecOps opencodeAgent (readFile sourceFile))
The calling software may start the tool.

These examples show fixed results. In use, the calling software maps actual tool calls to the actions in the rule set and asks the guard whether each action is permitted before calling the tool. It must block calls that return FORBIDDEN or UNDECIDABLE. Calls that bypass the guard are not checked.

03Checking the whole plan

Check the order
of a plan.

Even if build, test and deploy are each permitted, a plan violates the rule shown here when it puts deployment before the test of the artifact.

The agent or another program supplies the plan. Further plan rules cover the number of actions, time constraints and preconditions.

CI/CD · PLAN CHECKExample
01BuildbuildArtifact
02DeploydeployArtifact
03TesttestArtifact
The rule in MELD(obligateSequence
testArtifact deployArtifact)

Order violated. testArtifact must come before deployArtifact.

This example checks only the order. All other rules and preconditions must also be satisfied before the guard permits the plan.

04About the project

Articles on AgenticAEGIS

Visit the blog
MELD01

When AI Writes the Rules

Translating a domain requirement into a formal rule.

Read article
DDIC02

The Algorithm: How DDIC Resolves Conflicts

DDIC checks which of the conflicting rules takes precedence.

Read article
ARCHITECTURE03

Where the agent touches the wall

Checking calls in the executing software.

Read article

Research with open source code

Source code and examples

AgenticAEGIS is a research prototype. Its source code and the MELD files used in these examples are on GitHub.

Project on GitHub