Every day ships a runnable version
Start from an empty folder, verify each step with `uv run agent-code`, then add the next harness layer.
Build aCode AgentCLI in 14 days
Not a prompt collection. You will build the harness around the model: runtime, tool calls, permissions, session memory, and feedback loops.
Runnable Python + uv
Real Anthropic tool protocol
14 incremental snapshots
agent-code run
uv run agent-code
$ agent-code "read README, explain the CLI"
tool_use: read_file
observation: 42 lines
final: here is the harness boundary...
14
days
8
ready
1
CLI
The lesson is the harness
A code agent is not just prompts. It is tool boundaries, permission gates, context, execution, and feedback.
Docs, diffs, and replays together
Each lesson pairs prose with diffs, terminal traces, and Agent Loop diagrams so the system stays inspectable.
The first 7 days build one agent
Hello Agent, real model calls, file and web tools, safe edits, Bash permission, session memory, slash commands.
The next 7 days upgrade the harness
Plan Mode, Skills, Subagents, Context Compact, Coordinator, Worktree isolation, and tool authoring.
From one command to a working coding agent
The path is pain-driven: why providers exist, why tools need protocol shapes, and why file edits need diffs before writes.
1
Get started
Bring the CLI online and connect a real model.
2
Build
Add files, web, edits, Bash, and permissions.
3
Remember & control
Add session memory, slash commands, and hooks.
4
Scale the harness
Upgrade one agent into a planful, extensible system.
5
Ship & extend
Coordinate agents, isolate worktrees, and author tools.
Start at buildcc.dev and build the agent for real.
Day 1 is a tiny CLI that answers once. By Day 14, it can read code, edit files, run commands, and resume sessions.