Skip to main content

Overview

Lumie's agent system is a governed engineering surface. It is not a catalog of prompts, shortcuts, or one-off assistants. The .codex control plane defines how agent work is framed, routed, checked, reviewed, and reported across the workspace.

The market signal is clear: agent use is moving from experimentation into workflow infrastructure. McKinsey's 2025 State of AI survey reported broad AI adoption, active agent experimentation, and a gap between pilots and scaled enterprise value. Gartner's 2025 enterprise application forecast expects task-specific agents to become common in enterprise apps by 2026, while its June 2025 agentic AI forecast warns that many projects can fail when costs, business value, or risk controls are unclear. Deloitte's 2026 agentic AI research frames the same problem as agents scaling faster than guardrails, and PwC's 2025 AI Agent Survey points to workflow redesign and trust as the path beyond productivity gains.

Lumie's answer is to document agent engineering as an operating architecture: role-scoped agents, routed specialist work, human-in-the-loop ownership, evidence-producing reviewers, Tier 0 guardrails, and audit-ready workflow evidence.

What This Section Covers

PageUse it for
Agent ArchitectureAgent roles, ownership boundaries, routing layers, and bad patterns to avoid
Control Plane.codex surfaces, trust tiers, Dispatch Lock, hooks, lint, and checker boundaries
Workflow LoopsPlan, Work, Verify-Fix, Review-Fix, and Maintenance loop behavior
Review And EvidenceReview matrix coverage, evidence semantics, No findings, and commit readiness

Design Position

Lumie treats AI agents as infrastructure that needs clear ownership and verification. The system is designed around a few constraints:

  • agent output is evidence, not authority;
  • reviewer and checker agents do not replace required human-owner coverage;
  • orchestration stays in routed loops, not dispatcher agents;
  • rules that must hold without being read belong in hooks, lint, tests, or checkers;
  • specialist agents need narrow roles, explicit scopes, and reviewable output.

This keeps agent use practical. A main agent can ask a specialist for focused work or review evidence, but the main execution loop still owns Dispatch Lock, scope control, verification, and the final report.

Relationship To Development Docs

The Development section explains how to work in the Lumie workspace: repo layout, Tilt, and Docusaurus authoring. This section explains the agent operating model behind that work.

When a development task depends on entry gates, routing, agent review, or commit evidence, start here.

Market References