AI coding agents are general-purpose harnesses
These tools can do far more than write software—they're general-purpose agent harnesses.
The term "AI coding agent" is misleading. Tools like Droid, Claude Code, and Cursor are really just a language model, a set of tools (file system, shell, browser, APIs), and a loop. The "coding" part is just one use case. The real capability is autonomous tool use with reasoning.
The shell is the interface
What makes these tools powerful is that they can use the shell: curl for APIs, git for version control, ssh for remote machines, ffmpeg for media, python for scripting. If it can be done from a terminal, an AI agent can probably do it.
I use Droid for tasks that have nothing to do with code—research, data wrangling, system administration. It's my go-to for anything that involves multiple steps and reasoning through context.
Two patterns for agent context
MCP for custom tools. Model Context Protocol lets you expose any API as a tool the agent can use. I have MCP servers for my bookmarks API, Google Sheets, Google Calendar, and Perplexity search. This turns the agent into a personal automation layer. I can ask "how much did I spend on groceries last month?" and it pulls the data from my Tiller spreadsheet.
A caveat: MCP servers eat up context window quickly. Every tool definition consumes tokens before the agent starts working. Before adding another MCP server, ask: can the same thing be achieved with a shell script? Sometimes the answer is yes.
Markdown for memory. Factory.ai popularized a simpler pattern: markdown files that give agents persistent memory across sessions. AGENTS.md at your repo root for build commands and conventions. memories.md in ~/.factory/ for decisions and preferences.
Agents discover these automatically. This costs almost nothing in context and works across tools (Droid, Claude Code, Cursor, Gemini CLI). When you catch yourself explaining something twice, write it down. Within weeks you'll have a personalized assistant that knows exactly how you work. I use this with spec-driven development.
Why this matters
If you think of these tools as "AI coding agents," you'll use them to write code.
If you think of them as general-purpose automation tools, you'll use them for everything.
The second view is closer to reality—and way more useful.