Agency Coding Agent

Calling a large language model and asking it to "write some code" only gets you text. What turns that into an agent that can actually build software is the tool layer around it: the ability to read and modify real files, run commands to test what it just wrote, keep track of a multi-step plan, and loop until the work is done. Agency Coding Agent is a template that provides exactly that layer, built on the Agency Swarm framework with GPT-5.1 as the underlying model.

The agent is configured with OpenAI's Apply Patch tool for precise, diff-based file edits, a shell execution tool for running commands and tests inside a workspace directory, a plan-management tool for tracking multi-step tasks, an image generation tool for producing web assets on the fly, and web search for pulling in documentation or solutions mid-task. Together these let the agent scaffold a project, write and revise files, execute and observe the results, and adjust course — the same basic loop used by dedicated coding-agent products, exposed here as a template you can run yourself or extend.

A meaningful amount of the value is in the details most people underestimate: the shell tool auto-detects commands that would start a long-running dev server or watcher and forces them into the background instead of hanging the agent, applies non-interactive flags to common scaffolding commands (npm create, expo, django migrations, and similar) so the agent doesn't stall waiting on a prompt it can't answer, and enforces both a hard timeout and an inactivity timeout so a stuck process doesn't block the whole run. File edits are sandboxed to a workspace directory, with path resolution that rejects any operation that tries to escape it.

The repository ships as a Python project (Agency Swarm plus FastAPI) with a local terminal demo, a FastAPI entry point for serving the agent over HTTP, and a Dockerfile for containerized deployment, plus a deploy path to Agencii.ai for anyone who wants managed hosting rather than running it themselves. Because instructions, tools, and model settings are all defined in code, it's meant to be forked and customized — swap the system prompt, add a new tool, or point it at a different model — rather than used as-is.