Skip to content

Post-Plan Agent Swarm Proposal

Overview

After plan generation, PlanExe activates a specialized agent swarm. Each agent takes the plan artifact as input, does one job, and hands off. The GitHub repo (see plan-to-repo proposal) serves as the shared workspace — same role the codebase plays in Codebuff.

Agents

Repo Agent

Provisions a GitHub repo, pushes plan artifact, README, folder scaffold. (Ties into plan-to-repo proposal.)

Scaffold Agent

Reads plan phases and generates folder structure, basic website boilerplate, and project scaffolding appropriate to the plan type (SaaS, physical business, nonprofit, etc.)

Research Agent

Enriches the plan with live data: market size, competitor intel, pricing benchmarks, regulatory considerations. Commits findings to /research in the plan repo.

Issues Agent

Converts plan phases and tasks into GitHub issues and milestones. Creates an instant actionable task board on the repo.

Domain Agent

Suggests domain names aligned with the plan's business concept, checks availability, optionally registers (with user credit authorization).

Reviewer Agent

Critiques the plan for logical gaps, missing assumptions, and weak phases. Opens a GitHub issue or PR with suggested revisions.

Orchestration Pattern

  • Inspired by Codebuff's multi-agent architecture (File Picker → Planner → Editor → Reviewer)
  • Sequential where order matters (Repo Agent first), parallel where independent (Research + Issues + Domain can run simultaneously)
  • Each agent is stateless — reads from plan artifact + repo, writes back to repo
  • Model-per-agent: heavyweight model for Reviewer, lightweight for Issues/Domain

Credit Model

  • Basic swarm (Repo + Scaffold + Issues): included with plan generation
  • Research Agent: costs additional credits (external API calls)
  • Domain Agent: costs credits if registration triggered
  • Reviewer Agent: costs credits (heavyweight model inference)

Implementation Notes

  • Builds on plan-to-repo infrastructure
  • Agent definitions follow Codebuff-style TypeScript agent definition pattern
  • PlanExe orchestrates via job queue; agents are stateless workers