Architecture¶
This section provides an overview of the a5c platform architecture, describing its core components, configuration system, agent registry, and integration points. Together, these enable a Git-native, event-driven AI agent orchestration platform.
Overview¶
The a5c platform transforms a Git repository into an intelligent, autonomous development environment. Agents are defined as files in the repository and are triggered by GitHub events (e.g., commits, pull requests, issue comments) via a GitHub Action. The system components load resources, manage configuration, route events to agents, execute prompts, and process structured outputs.
+----------------------+ +--------------------------+ +-------------------+
| GitHub Repository | Git | a5c GitHub Action | | Model Context |
| (code + .a5c/agents) | <----> | (.github/workflows/a5c.yml) |---->| Protocol Servers |
+----------------------+ +--------------------------+ +-------------------+
| | |
v v v
+---------------------------+
| Core Components |
| - Resource Handler |
| - Agent Router |
| - Config Manager |
| - Prompt Engine |
| - Agent Execution |
| - Output Processor |
+---------------------------+
|
v
+---------------------------+
| AI Agents (Registry) |
+---------------------------+
Core Components¶
The GitHub Action implementation (see a5c-ai/action) consists of focused modules:
Component |
Responsibility |
|---|---|
Resource Handler |
Load prompts, agents, and configuration from local or remote sources with caching and retry logic. |
Agent Router |
Discover agents based on events, mentions, labels, and other triggers. |
Config Manager |
Merge built-in defaults, user configuration ( |
Prompt Engine |
Render templated prompts using YAML frontmatter and context. |
Agent Execution |
Invoke the configured LLM or CLI tool and collect the raw response. |
Output Processor |
Parse and format structured data from agent responses. |
MCP Manager |
Manage built-in Model Context Protocol (MCP) servers (e.g., filesystem, memory, search, GitHub). |
Utilities |
Common helpers, logging, and error handling. |
Configuration System¶
The configuration hierarchy enables flexible overrides and remote loading:
Built-in Defaults (
default-config.yml)User Configuration (
.a5c/config.ymlor remote URI)Agent Frontmatter (YAML in
.agent.mdfiles)
Configuration files and agents can be sourced from remote repositories or URLs, authenticated via GITHUB_TOKEN. See Configuration · a5c-ai/action.
Specification Repository¶
The platform specifications are maintained in the a5c-ai/spec repository. This includes the Agent Format and Proposal System specifications:
Agent Format Specification: Defines YAML frontmatter structure, inheritance model, and validation rules.
Proposal System Specification: Outlines the community governance and proposal workflow.
Agent Registry¶
A community-driven registry (see a5c-ai/registry) hosts shared agents, enabling composable intelligence through inheritance. Base agents published in the registry can be extended locally or remotely.
Seed Templates¶
Starter templates for new a5c projects are available in the a5c-ai/seed-generic repository. Use these seeds to bootstrap custom projects with best-practice frontmatter and prompts.
Integration Points¶
GitHub Actions: Event-driven invocation via
.github/workflows/a5c.yml.CLI Tools: Supports multiple LLMs (e.g., GPT, Claude, Gemini) through a unified CLI interface.
Model Context Protocol: Integrates MCP servers for enhanced context (e.g., search index, GitHub API).