Skip to main content
jeff.leung
Case study 02 · Updated 2026-05-27

Agent Tools Workstation System

A reproducible AI engineering environment — scripted setup, shared skills, sub-agent personas, host integrations, and conventions that make a single workstation behave like a small platform.

AI-assisted engineering Developer experience Infrastructure Backend / Platform
Problem

Coding-agent productivity collapses when the engineering environment drifts. Skills, MCP servers, sub-agent personas, hook configuration, and host wiring break in subtle ways across hosts and machines. Without a reproducible workstation, AI-assisted reliability becomes a story about a specific laptop.

Constraints
  • Must work across Codex CLI, Claude Code, OpenCode, Cursor, and other coding-agent hosts using shared conventions.
  • Must keep secrets out of the version-controlled layer.
  • Must support strict-idempotent installs so onboarding a new machine is reproducible, not heroic.
  • Must let one person maintain it without a team-sized DevOps cost.
Architecture

The agent-tools workstation system is what stops AI-assisted engineering from being a story about one configured laptop. It packages the skills, sub-agent personas, host integration adapters, and MCP server wiring into a single installable repo. When the installer runs, the workstation converges to a known good state.

The strict-idempotent install

Running setup twice produces no diffs the second time. That sounds boring; in practice it is the property that makes onboarding a new machine survivable. The installer is split into atomic steps. Each step is a small unit that can be retried, skipped on convergence, or rolled back. A doctor flow inspects the workstation and proposes the next safe step.

The skill library and sub-agent personas

Skills are reusable instructions an agent can follow. Sub-agents are personas with focused tools and constrained scopes. Both live in one repository and are linked into each host’s expected location, so a skill written once is usable from any supported coding-agent host. This keeps reliability work portable.

What the workstation is not

It is not a team platform. It does not aspire to be one. The design optimizes for a single engineer who needs a reliable AI engineering environment across multiple hosts, not for cross-team governance. The constraints are honest, and the system is sized to them.

Reliability mechanisms
  • Strict-idempotent install — running setup twice converges to the same state; partial failures are recoverable.
  • Single source of truth — skills, sub-agent personas, and conventions live in one repo and are linked into each host's expected location.
  • Host-aware adapters — each coding-agent host gets a thin integration that surfaces shared skills without forking content.
  • Local-first — no centralized server is required; the workstation is the platform.
  • Doctor flow — `status / plan / exec / health / verify / report` to diagnose a drifted workstation and bring it back to convergence.
Tradeoffs
  • Single-engineer maintenance means no SRE; reliability comes from convention discipline, not from staffing.
  • Cross-host parity requires occasional rewrites when a host changes its plugin model.
  • Reproducibility is purchased with a strict directory and configuration contract, which constrains personal tweaks.
What this proves

Demonstrates the ability to treat a personal AI engineering environment as a small platform — one with reproducible install, shared conventions, and a doctor flow — which is what makes the rest of the reliability story possible on more than one machine.

Privacy notes

No host vendor lock-in is claimed; references to coding-agent hosts use only their public product names. No internal tooling, organization, or repository is named.