Skip to content

Multi-agent strategy

The project goal is practical:

Make Plus feel like Pro.

Use the strongest model for judgment. Use cheaper or faster workers for scoped execution. Keep Codex in charge.

Operating model

LayerRole
BrainCodex plans, routes, reviews, and decides
HandsClaude Code workers perform scoped tasks
RouterCCSwitch provides local profiles and models
LedgerEach run leaves metadata, prompts, stdout, and stderr

Default roles

RoleUse it for
requirementsScope, non-goals, acceptance criteria
architectureRepo map, likely files, implementation plan, risk
developmentMain code development tasks
testingTest plan, edge cases, validation commands
reviewFindings, file references, maintainability
performanceRuntime, IO, latency, resource use
compatibilityWindows, macOS, Linux, shell, version risk
documentationTutorials, FAQ, examples, onboarding
automationCI, release workflows, package checks
securitySecrets, permissions, command risk, supply chain
implementationScoped edits when write access is allowed
opsDeployment, logs, rollback, runtime risk
multimodalImage or mixed input work

Four-phase workflow

  1. Parallel analysis: requirements, architecture, security, testing, and other relevant roles inspect the task.
  2. Cross-review: agents compare risk, scope, and plan quality.
  3. Execution: implementation runs only after the plan is stable and write scope is clear.
  4. Controller summary: Codex reviews logs, diffs, tests, and final output.

Generate the plan:

bash
python "$CC_ORCHESTRATOR_HOME/cc_orchestrator.py" workflow-plan "Ship a safe refactor"

Cost rule

Do not spend the best model on every subtask.

Use it for:

  • final judgment
  • architecture
  • risky review
  • hard tradeoffs

Use worker models for:

  • repo mapping
  • test design
  • documentation draft
  • compatibility checks
  • focused implementation after approval

This is the main reason the project exists.

MIT licensed. Not affiliated with OpenAI, Anthropic, Claude, Claude Code, or CCSwitch.