Easy Agent

An open-source, terminal-native agentic coding CLI (TypeScript / Node.js, MIT) that recreates a Claude Code-style workflow — built less as a product than as a teaching scaffold: “want to understand coding agents? rebuild one stage at a time.” It makes the agentic-coding-harness‘s anatomy explicit instead of hiding it behind one wrapper. First-party repo (ConardLi), surfaced via a Dan Kornas tweet from the Telegram channel. URL-only ingest.

What it is

A working coding agent split into five named layers you can read independently: provider I/O (model communication), tooling, the agent loop, session/orchestration flow, and the terminal UI. It ships a React/Ink terminal interface (a local CLI experience, not a web app), a real tool surface (file, shell, search, web, MCP, local actions, plus permissions and sandboxing), and multi-provider support (Anthropic, OpenAI-compatible, Gemini, and local OpenAI-compatible endpoints like Ollama). Its distinctive move is 31 step-by-step snapshots: completed stages 1–31 each have a standalone step file you study alongside the main tree, so you can watch the harness accrete one capability at a time.

Why it belongs here

• The harness, dismantled into this spoke’s own layers. The synthesis names the value as “everything wrapped around the model” — provider plumbing, tools, the loop, orchestration, durability. Easy Agent is that thesis turned into a curriculum: it literally separates harness into provider I/O + tools + loop + session + UI, the same decomposition LangChain frames as agent = model + harness (langchain-custom-harness). For a reader, it’s the cleanest way to see that the “magic” is engineering — the layers, made legible.
• The augment/teaching pole, aimed at agent-building itself. It rhymes with lathe‘s “LLMs to teach you, rather than think for you” — but where Lathe teaches arbitrary skills by hand, Easy Agent teaches how a coding agent is constructed by rebuilding one. A pedagogical sibling to the working harnesses (claw-code, oh-my-pi): same parts, optimized for study over throughput.
• Provider-agnostic by construction. Like claw-code and tgpt, it treats the model as a swappable backend (four provider families incl. local Ollama), reinforcing that the harness — not any one model — is the artifact.

Caveat

A learning-oriented reference implementation, not a benchmarked or production harness — its value is legibility, not performance, and the “31 stages” are a snapshot of an in-progress build. Author (ConardLi) noted inline rather than as an entity node — GitHub handle, low graph signal, matching the skill-pack/harness precedent.

Related

agentic-coding-harness · claw-code · oh-my-pi · lathe · langchain-custom-harness · synthesis