Silly Agentic Coding 🤖

A repository dedicated to defining, organizing, and coordinating specialized AI agents. This workspace serves as a template directory and coordinate registry for agentic coding workflows, establishing standard personas, boundaries, and rules of engagement to preserve code quality and prevent operational conflicts.

🚀 Concept

In modern AI-assisted software engineering, having a single model manage all concerns can lead to context pollution, scope creep, and conflicting edits. This repository implements a multi-agent orchestration framework by dividing development concerns into narrow, highly specialized roles (e.g., performance tuning, testing, security, Kubernetes configuration, and pipeline management).

Each agent is governed by:

• A specific persona and target mission.
• Core boundaries (things they must always do, ask first, or never do).
• A daily execution process.
• An individual journal (.jules/<agent-name>.md) for recording critical environment or syntax discoveries.

---

🗂️ Agent Categories

All agents are documented in the AI Agents Directory (AGENTS.md). Here is the category breakdown:

1. Architecture & Code Quality: Architect, Steward, Gardener, Curator
2. User Experience & Design: Navigator, Palette, Quill
3. Testing, Security & Resilience: Inspector, Sentinel, Sledge, Bolt
4. Kubernetes & Infrastructure Configs: Operator, Registrar, Helmsman, Tailor
5. CI/CD & Delivery Pipelines: Butler, Mason, Stagehand
6. Observability, Telemetry & Onboarding: Beacon, Watcher, Compass, Scribe, Pathfinder

---

📋 Rules of Engagement

To keep the workspace clean, all agents are expected to adhere to the following rules defined in AGENTS.md:

1. Strict Scope Separation: Respect each agent's domain. Never edit files or configurations belonging to another agent's scope without explicit authorization.
2. Atomic Contributions: Keep changes focused and small (ideally under 50 lines of code changes).
3. Mandatory Local Verification (FullThrottle): Every agent must pass the Pre-PR Verification Gate (test → rebase → retry loop) before submitting any PR. See agent_sources/common/pr_gate.md.
4. Zero-Trust Security: Never commit raw secrets or credentials. Always utilize secret injection libraries or environment configuration variables.
5. Continuous Learning: Keep logs concise and focused strictly on critical environment, framework, or tooling learnings in .jules/<agent-name>.md.

---

🛠️ Development & Building Prompts

This project uses a modular prompt architecture. Agent prompts are constructed from reusable components under agent_sources/common/ and agent-specific templates under agent_sources/templates/ (standard) and agent_sources/bulk_templates/ (bulk refactoring).

To compile all monolithic output agent files into agents/ and agents_bulk/:

python3 build.py

To verify that the compiled agent files are up-to-date with their sources (e.g., in CI/CD or pre-commit hooks):

python3 build.py --check

Warning

Never edit the files inside the agents/ or agents_bulk/ directories directly. They are auto-generated and will be overwritten by build.py. Make all modifications under agent_sources/ instead.

🚀 Usage

When invoking an agent, mount its compiled monolithic prompt markdown file as a system prompt instruction:

• For daily autonomous repository maintenance and small cleanups: use agents/<agent-name>.md (e.g., steward.md).
• For targeted multi-file deep refactoring and code health tasks: use agents_bulk/<agent-name>.md (e.g., steward.md).