Building Portable Agent Control Planes

Resources

• AGENTS.md — The format’s own site describes repository-local instructions for coding agents, including nested guidance. Use it as a small, inspectable foundation for portable project knowledge while keeping harness-specific execution and permission behavior explicit. https://agents.md/

• Agent Skills Specification — Defines the skill directory, Markdown frontmatter, supporting resources, and progressive disclosure. Provides a concrete portability contract for reusable capabilities without implying that every harness implements identical tools or invocation semantics. https://agentskills.io/specification

• Claude Code Skills — Shows executable Markdown in practice through skill instructions, dynamic context injection, supporting scripts, and invocation controls. Useful for separating a reusable workflow’s intent from the Claude-specific adapter features needed to execute it. https://code.claude.com/docs/en/skills

• Claude Code Memory and Project Rules — Documents project instructions, imports, and scoped rules. Use it to design contextual loading and documentation delivery, then identify which paths and loading conventions must change when migrating to another harness. https://code.claude.com/docs/en/memory

• Claude Code Settings and Precedence — Describes configuration scopes and precedence for an agent environment. A useful concrete reference for mapping profiles, permissions, hooks, and tool configuration into an inspectable control plane rather than scattering settings across machines. https://code.claude.com/docs/en/settings

• GitHub Spec Kit — An open-source specification-driven workflow toolkit with integrations for multiple coding agents. Study its shared templates and agent-specific integration structure as an example of versioned workflow definitions with explicit harness adapters, not proof of identical behavior across tools. https://github.com/github/spec-kit

• Agent Client Protocol — Defines a protocol between coding agents and their clients. Helps separate the editor-facing adapter boundary from the durable workflow, agent roster, and policy configuration that a portable control plane still needs to own. https://agentclientprotocol.com/get-started/introduction

• Model Context Protocol Architecture — Explains hosts, clients, servers, and negotiated capabilities for tools and context. Useful for making tool connectivity replaceable while recognizing that instruction loading, authorization, and workflow state remain separate concerns. https://modelcontextprotocol.io/docs/learn/architecture

• Development Container Metadata — The specification documents reproducible environment configuration, lifecycle commands, mounts, and environment variables. Use it to make runtime assumptions and setup shims explicit; a reproducible container is not by itself a complete sandbox or secret-management policy. https://containers.dev/implementors/json_reference/

• Claude Code Sandboxing — Explains filesystem and network isolation, platform requirements, and sandbox configuration. Provides a concrete checklist for preserving safe defaults and credential boundaries when moving workflows between local and remote environments. https://code.claude.com/docs/en/sandboxing

• Aider’s Repository Map — Describes a compact, relevance-ranked map of a codebase that is supplied to the model. Useful for designing task-aware knowledge delivery and context budgets instead of expecting every agent to rediscover architecture through file search. https://aider.chat/docs/repomap.html

• Git Reflog — Git’s reference for inspecting prior positions of local references and recovering otherwise hard-to-find commits. Ground control-plane recovery in committed configuration and inspectable history, while noting that reflogs are local, expire, and are not a substitute for shared backups. https://git-scm.com/docs/git-reflog