System Design

Agentic Systems Architecture

How Claude and Codex operate as complementary runtimes over shared MCP integrations, CLI tooling, quality gates, and the local development environment.

Agentic Systems Map

One architecture, multiple agents, shared execution fabric

This model reframes the stack around runtime responsibilities rather than a single assistant. Claude and Codex operate as complementary runtimes over the same protocol, tooling, and deployment surfaces.

2

Runtime Profiles

4

Architecture Layers

4

Execution Flows

2

Evidence Sources

Runtime Profile

Claude Runtime

Judgment-heavy orchestration and workflow policy

Coordinates planning depth, skill routing, fallback strategy, and quality gates across complex tasks.

  • Modular playbook with integrations, standards, and workflow docs
  • Explicit fallback chain for protected web automation
  • Delegates mechanical work to Codex when parallelism matters

Runtime Profile

Codex Runtime

Execution-oriented implementation and tooling throughput

Runs high-effort coding sessions with MCP-backed tools, trusted project scopes, and fast terminal execution.

  • Primary model profile set to gpt-5.3-codex with xhigh reasoning
  • Configured MCP server registry for browser, infra, and productivity systems
  • Unified exec and shell snapshots enabled for repeatable command workflows

Layered Systems Diagram

Layer Topology

The orchestrator coordinates four durable layers. Each layer is independently evolvable but connected through explicit interfaces.

Core Hub

Agent Runtime
Orchestrator

Agent Runtimes

Top-layer reasoning engines that decide intent, delegation boundaries, and execution strategy.

Claude CodeCodex CLIHandoff ProtocolsSub-agent Mesh

Protocol & Tooling Fabric

Shared integration layer that normalizes external capabilities into stable interfaces.

MCP ServersSkill GatewaysCLI BridgesLSP Services

Execution Surfaces

Operational targets where changes are created, verified, and shipped.

Repository WorkspaceGit + Quality GatesPlaywright E2EVercel / Cloudflare

Host Environment

Local workstation stack that anchors terminal, editor, and review ergonomics.

GhosttyZedLazyGit + GitHub DesktopmacOS Workstation

Flow 1

Intent Routing

Prompt -> Runtime Selection -> Task Shape

Requests are classified first, then directed to the runtime best suited for judgment, speed, or scale.

Flow 2

Capability Invocation

Runtime -> MCP/Skills/CLI -> External System

The protocol fabric provides consistent adapters so agents can invoke tools without custom glue per task.

Flow 3

Verification Loop

Code Change -> Lint/Test -> Visual/E2E

Every significant edit flows through deterministic checks before being considered production-ready.

Flow 4

Delivery Loop

Commit -> Deploy -> Runtime Feedback

Deployment and observability signals feed back into both runtimes to tighten future execution decisions.

Configuration Evidence

Claude Configuration Evidence

Source: ~/.claude/claude.md + ~/.claude/docs/integrations.md

Curated from orchestration, fallback, and integration guidance; secrets and sensitive values removed.

  • Modular instruction system with dedicated integration and workflow docs
  • Skill-first workflows with explicit Codex delegation pathways
  • Integration policy that prioritizes structured APIs before browser scraping

Codex Configuration Evidence

Source: ~/.codex/config.toml + ~/.codex/instructions.md + repo .codex

Summarizes model/runtime defaults, MCP inventory, and execution features from local Codex setup.

  • Default runtime tuned for high reasoning depth on coding tasks
  • Configured MCP set spans browser automation, infra deployment, and life tooling
  • Project trust model and unified execution features enabled