Terminology & Architecture
Multi-Agent Systems Glossary
Precise, plain-language definitions for multi-agent orchestration, independent verification, private fleet architecture, and digital sovereignty.
Architecture & Infrastructure
Fleet
A coordinated pool of user-owned workstations, laptops, servers, and GPU instances executing agent tasks under a unified control daemon.
Harness
The runtime execution environment and multiplexer that hosts an agent, injects tools, manages terminal input and output, and logs telemetry.
Agent
An autonomous software process combining a model reasoning loop, tool execution permissions, and workspace state to achieve specified objectives.
Git Worktree
An isolated filesystem directory linked to a Git repository, allowing parallel agents to modify branches independently without workspace collisions.
Digital Sovereignty
The principle that users own their compute hardware, private data, code repositories, and model weights without landlord lock-in.
On-Device Context
Sensitive codebase context, syntax trees, and documentation indexed and maintained locally on host hardware rather than uploaded to vendor clouds.
Private Appliance
A planned dedicated on-premise hardware deployment hosting pre-quantized models with local data boundary isolation.
Rate Limit Multiplexing
Routing parallel subagent tasks to spread work across machines and providers you are entitled to use.
Verification & Security
Verifier
An independent agent running a distinct model family tasked with auditing, executing test suites, and confirming work before a task is marked complete.
Rework
The automated cycle wherein a verifier rejects a diff, sending detailed failure diagnostics back to the builder agent for corrective iteration.
Cross-Vendor Verification
A consensus mechanism where artifacts created by one model provider are evaluated and verified by a competing model family.
AST Diff Analysis
Abstract Syntax Tree inspection comparing semantic code modifications rather than superficial whitespace or formatting changes.
Continuous Adversarial Exposure Validation
Automated, defensive testing to discover exploitable security flaws in authorized systems through continuous probes.
Deterministic Vulnerability Verification
Confirming a security flaw by constructing an executable, repeatable reproducer script that validates the weakness.
Security Control Efficacy Testing
Measuring whether defensive tools, firewalls, and detection rules successfully alert on or block candidate attack vectors.
Continuous Patch and Remediation Testing
Re-running verification probes against candidate patches to confirm vulnerability mitigation without functional regressions.
Zero Data Retention (ZDR)
An architectural and policy commitment ensuring client code, prompts, and outputs are not retained on shared infrastructure after processing.
Models & Authentication
Local Model
An open-weight model executing entirely on local host hardware using local inference runtimes without sending tokens across external networks.
OAuth 2.1
The authorization framework using Authorization Code flow with PKCE, exact redirect URI matching, and refresh token rotation to secure agent connectors.
PKCE (Proof Key for Code Exchange)
A cryptographic protocol extension to OAuth that prevents authorization code interception attacks on native applications and agents.
Bring Your Own Key (BYOK)
An architectural pattern where customers supply their own vendor API keys directly, retaining billing control and direct provider relationships.
Client ID Metadata Document
A URL-hosted JSON specification allowing OAuth 2.1 clients to declare client identity, logo, and scopes without manual portal registration.
Protocols & Sandboxing
Model Context Protocol (MCP)
An open standard protocol enabling agents to discover, authenticate, and query external tools and data sources through structured interfaces.
Agent-to-Agent (A2A)
Direct machine-to-machine communication protocols enabling autonomous agents to negotiate sub-tasks, delegate work, and exchange artifacts.
Agent Client Protocol (ACP)
A standardized protocol allowing editors, harnesses, and client interfaces to communicate with agents, manage handoffs, and negotiate approvals.
MicroVM Sandbox
A lightweight hardware-isolated virtual machine providing near-instant spin-up and strict kernel boundaries for untrusted code execution.
Session Forking
Branching an active agent execution state, filesystem snapshot, and context history into parallel speculative trajectories.
Monetization & Governance
Aura by OmniConflux
A planned privacy-first contextual advertising network for OmniConflux surfaces, designed around on-device topic classification. Not live yet.
Contextual Ad Broker
A local and central service designed to run parallel expected-eCPM auctions for developer-relevant placements without sharing code or prompts.
Learn More
Explore the System Architecture
Discover how these components operate together in our technical documentation, fleet topology guides, and verification specifications.