Agentic Software Engineering Workshop

Overview

This directory contains the following files:

• compose.yaml - Docker Compose file to launch OpenCode container. Inspect it before launching the container.
• opencode.jsonc - OpenCode configuration file. Update it as needed, especially to add your specific provider and model configurations. Also see OpenCode config documentation.
• agents/ - Agent definitions and permissions. See OpenCode agents documentation.
• workspace/ - A directory containing sample project artifacts based on a previously vibe-coded project. See README.md and proposal.md for more details. Replace them if you have another project in mind.

Tip

Ask your friends with self-hosted LLMs for their OpenCode configuration files.

Getting Started

Install Prerequisites

On Windows, Install Docker, git, and WSL. Docker will automatically use WSL as its backend.

# install git and docker in your Windows environment, not in WSL
winget install Git.Git Docker.DockerDesktop

# Verify that you have WSL 2
wsl --version

# if not present, run `wsl` in PowerShell as admin to install
wsl

# if already installed, try updating
wsl --update

# reboot if needed

After installing Docker and WSL, open Docker Desktop and start the Docker Engine. This uses WSL as a backend.

Note

On any other OS, just ensure that you have git and docker available.

Obtain the working directory

Clone this repository.

git clone ...
cd ...

Inspect the contents.

• compose.yaml: docker compose service specification
• opencode.jsonc: OpenCode configuration; update as needed
• workspace: your working directory
• agents: system prompts / agent roles provided to you; create more of them!

Launch OpenCode

docker compose up

To run OpenCode on the command line, try the following:

docker compose run --help
docker compose run stats
docker compose run workspace

Tip

To stop a detached container, run docker compose down --remove-orphans. Check container status by running docker compose ps in the same directory or docker ps anywhere.

Open http://localhost:4096 in your browser.

Configure OpenCode

Under OpenCode Settings

• Show Reasoning Summaries
• Expand shell tool parts
• Advanced > Show agent

Add providers. Mistral and Poolside provide free access. Opt out of data collection if the provider allows. OpenRouter and OpenCode Zen also provide access to free and paid models. Ultimately, consider hosting local models.

Tip

Most self-hosted providers provide an OpenAI-compatible API endpoint. OpenCode allows you to add any OpenAI-compatible provider from the UI. These URLs typically end in /v1. To get a list of models available from your provider, open the provider URL in a browser and append /v1/models to the end of the URL.

Warning

Many free models available by default in OpenCode retain and use your data for training. Also ensure that you are authorized to use the specific models beforehand. Use the Manage Models setting to disable models that you do not intend to use.

Add the project folder ~/workspace to your OpenCode interface. You can also mount and use other folders as needed- see compose.yaml.

Create a new session and start interacting with various agents. Select your preferred model. Each agent can use a different model. Some agents have more permissions than others. See OpenCode permissions documentation and update agent definition files in agents directory as needed.

Tip

You can create new agents to define your typical workflows by running docker compose run agent create

Proposed Agentic Workflow

flowchart LR
    Proposal --Architect --> Design
    Proposal --Analyst --> Specifications
    Specifications --Architect --> Design
    Specifications --Planner --> Tasks
    Design --Planner --> Tasks

Loading

Artifacts

• Proposal describes the problem, scope, and user intent. It answers the why and what.
• Specifications include functional and non-functional requirements, constraints, and risks. It answers the what and how.
• Design describes the system architecture, data flow, interfaces, and dependencies. It answers the how.
• Tasks are a concrete, ordered implementation checklist -- backlog of work items with clear completion criteria.

Agents

• Analyst: Translate user intent and proposal into a clearly defined set of specifications including functional and non-functional requirements, constraints, and risks
• Architect: Convert proposal and specifications into a workable technical design
• Planner: Convert design and specifications into an executable implementation plan
• Technical Manager: oversee project execution; chunk
• Developer: Implement code, run commands, write unit tests, run lint/typecheck, investigate errors, and debug
• Reviewer: Review correctness, clarity, maintainability, and requirement alignment
• Tester: Design and execute unit, integration and end-to-end tests; validate test quality and coverage of required behavior

Tip

A special Minion agent is provided to you with a minimal system prompt for tasks that do not require extensive instruction.

References

• Talk: Agentic Software Engineering
• OpenCode
• Mistral Small 4 / Medium 3.5
• Poolside Laguna XS 2.1