Agent Architecture

Agents are role-scoped adapters behind Lumie's routing system. They are not a dispatcher layer, a replacement for the main execution loop, or an approval mechanism. The main agent still owns Dispatch Lock, loop state, scope control, and the final report.

Role Model

Role	Responsibility	Boundary
Main agent	Frames the task, resolves owner surface, applies the accepted loop, integrates evidence, and reports outcome	Does not bypass Plan Mode, hooks, checks, or human-owner coverage
Specialist writer	Makes bounded edits inside an explicit repo, module, or file scope	Does not own unrelated files or revert user changes
Reviewer agent	Reads a diff or document surface and returns severity-ordered findings	Produces evidence; does not approve the work
Checker agent	Produces policy, risk, security, contract, or consistency evidence	Cannot replace tests, hooks, or human-owner review
Human owner	Holds mandatory review ownership for the changed surface	Required owner coverage is not satisfied by agent evidence

This model keeps agents useful without turning them into an unreviewable workflow harness.

Routing Layers

Agent selection flows through three routing layers:

1. .codex/AGENTS.md establishes the entry gate, owner surface, Dispatch Lock, and whether a specialist is needed.
2. .codex/routing/repos/*.md maps product repo work to writer guidance, verifier guidance, and repo-specific rules.
3. .codex/routing/review-matrix.md maps the actual diff to required human owners, reviewer agents, checker agents, and Verify-phase evidence.

The output of this routing is an evidence plan. It does not grant permission to skip verification or ownership.

Agent Families

Family	Purpose	Examples
Writers and scaffolders	Make bounded edits inside a routed write scope	frontend-scaffolder, backend-scaffolder, worker-scaffolder, lumie-doc-author, codex-md-author
Reviewers	Read a diff or document surface and return findings	frontend-reviewer, backend-reviewer, worker-reviewer, lumie-doc-reviewer, lumie-doc-i18n-reviewer
Checkers and tripwires	Produce policy, risk, security, or contract evidence	security-tripwire, anti-pattern-tripwire, api-contract-checker, lumie-risk-scorer
Verification specialists	Run or summarize affected checks for a repo surface	frontend-test-orchestrator, backend-test-orchestrator, worker-test-orchestrator, performance monitors
Explorers and domain specialists	Answer navigation, terminology, or architecture questions without editing	codebase-explorer, infra-explorer, lumie-domain-expert, lumie-architect
Curators	Maintain rules, references, and learning artifacts	reference-curator, lumie-rulebook-curator, lumie-rulebook-pruner, post-mortem-author

Write-capable agents need an explicit ownership slice. Read-only agents should return severity-ordered findings or a clear No findings result. Both kinds must work with existing user changes instead of reverting unrelated work.

Evidence, Not Authority

The important boundary is that agent output is evidence, not authority. A reviewer or checker can produce findings, confirm no findings, or provide a specialized signal. It cannot replace a required human-owner:* row, approve a commit, or bypass a hook, lint, test, or checker gate.

This is why Lumie avoids dispatcher agents and workspace workflow harnesses. The architecture keeps orchestration in the routed loop, keeps specialist behavior small and inspectable, and keeps enforcement in Tier 0 mechanisms when a rule must hold without human reading.

Patterns To Avoid

Pattern	Why it is rejected
Dispatcher agents	They hide routing decisions behind another prompt layer and make ownership harder to audit
Agent sprawl	Too many overlapping specialists make review routing ambiguous
Slash-command harnesses	They turn workflow control into invocation ceremony instead of observed loop state
Reviewer-as-approver	It confuses evidence with authority and can bypass human-owner coverage
Document-only guardrails	Critical rules fail when they depend only on someone reading prose

Related Pages

• Control Plane for the trust-tier and enforcement model.
• Workflow Loops for the execution loop that routes agents.
• Review And Evidence for reviewer, checker, and human-owner evidence semantics.