docs(blog): local agents on AMD GPUs, with runnable examples (#14)

2026-07-21 20:20:51 +08:00
parent 0e134e9c6c
commit d5c2bc001d
10 changed files with 1095 additions and 0 deletions
@@ -0,0 +1,75 @@
+<!-- English | [简体中文](README.zh.md) -->
+
+# Example: an Agent that builds another Agent (local, on an AMD GPU via Ollama)
+
+This example is the **"Harness for Building Agents"** pillar in runnable code. Using only the
+PenguinHarness SDK, it:
+
+1. **builds** a brand-new agent (`commit-helper`) by driving `default_agent` with the
+   `agent-creation` skill from a plain-language requirement, then
+2. **runs** that freshly-created agent to prove the generated `AGENTS.md` actually shapes its
+   behavior (it writes a Conventional Commits message).
+
+Everything runs on a **local open-weight model** — `qwen3.6:35b` served by Ollama — so no cloud
+API and no data leaving the machine. Ollama's ROCm backend runs this natively on AMD GPUs
+(from Radeon PRO workstation cards up to Instinct accelerators).
+
+## 1. Serve the model locally with Ollama
+
+```bash
+# Ollama detects an AMD GPU (ROCm) automatically; pin a specific card if you like:
+export HIP_VISIBLE_DEVICES=0
+ollama serve &          # if not already running as a service
+ollama pull qwen3.6:35b
+```
+
+## 2. Point PenguinHarness at it (once)
+
+```bash
+penguin config model add \
+  --model-id qwen3.6:35b \
+  --provider custom --client-type openai \
+  --base-url http://localhost:11434/v1 \
+  --api-key ollama --set-default
+```
+
+This writes the model into `~/.penguin/data/default_project/.project_config.toml`. The example
+uses the project's default model, so no model id is hard-coded in the script.
+
+## 3. Run the example
+
+From the repo root (build the workspace first so `@prismshadow/penguin-core` resolves to its
+`dist/`):
+
+```bash
+pnpm install
+pnpm build
+pnpm --dir examples/build-agent-with-agent start
+# or directly:  npx tsx examples/build-agent-with-agent/build-agent.ts
+```
+
+## What you should see
+
+- **Phase 1** — `default_agent` scaffolds `agents/commit-helper/` under the project: its
+  directory layout, a copied `system_config.yaml` (with name + description), and an `AGENTS.md`
+  encoding the Conventional Commits rules.
+- **Phase 2** — the new `commit-helper` agent, following only that generated `AGENTS.md`,
+  produces something like:
+
+  ```text
+  fix(payment): add retry-with-backoff for transient gateway 503 errors
+
+  Transient 503 responses from the payment gateway were causing checkout
+  failures during peak traffic. Retry with exponential backoff gives the
+  gateway time to recover, preventing spurious user-facing errors.
+  ```
+
+## Notes
+
+- Output quality tracks the model. `qwen3.6:35b` handles this task well; much smaller models
+  may not follow the tool-calling protocol reliably.
+- A capable model occasionally introduces a mechanical slip when re-serializing the base config
+  (e.g. an invalid YAML escape) — every file is plain text and every run is traced, so such
+  slips are quick to spot and fix. This is the realistic shape of "agents building agents":
+  the requirement→`AGENTS.md` heavy lifting is automated; a human reviews the mechanical edges.
+- Re-running Phase 1 will update the existing `commit-helper` agent in place.
@@ -0,0 +1,70 @@
+<!-- [English](README.md) | 简体中文 -->
+
+# 示例：用一个 Agent 构建另一个 Agent（本地，经 Ollama 跑在 AMD GPU 上）
+
+这个示例是 **“构建 Agent 的 Harness”** 支柱的可运行版本。仅用 PenguinHarness SDK，它会：
+
+1. **构建**一个全新的 Agent（`commit-helper`）——用 `agent-creation` skill 驱动 `default_agent`，
+   根据一句大白话需求把它搭建出来；然后
+2. **运行**这个刚创建的 Agent，证明生成的 `AGENTS.md` 确实塑造了它的行为（它会写出一条
+   Conventional Commits 提交信息）。
+
+全程跑在一个**本地开源权重模型**上——由 Ollama 提供的 `qwen3.6:35b`——所以不用任何云端 API、
+数据也不离开本机。Ollama 的 ROCm 后端能原生在 AMD GPU 上运行它（从 Radeon PRO 工作站显卡一直到
+Instinct 加速卡）。
+
+## 1. 用 Ollama 在本地提供模型
+
+```bash
+# Ollama 会自动识别 AMD GPU（ROCm）；也可以指定某张卡：
+export HIP_VISIBLE_DEVICES=0
+ollama serve &          # 若尚未作为服务运行
+ollama pull qwen3.6:35b
+```
+
+## 2. 把 PenguinHarness 指向它（一次即可）
+
+```bash
+penguin config model add \
+  --model-id qwen3.6:35b \
+  --provider custom --client-type openai \
+  --base-url http://localhost:11434/v1 \
+  --api-key ollama --set-default
+```
+
+这会把模型写进 `~/.penguin/data/default_project/.project_config.toml`。示例使用项目的默认模型，
+所以脚本里没有硬编码任何 model id。
+
+## 3. 运行示例
+
+从仓库根目录（先构建 workspace，好让 `@prismshadow/penguin-core` 解析到它的 `dist/`）：
+
+```bash
+pnpm install
+pnpm build
+pnpm --dir examples/build-agent-with-agent start
+# 或直接运行：  npx tsx examples/build-agent-with-agent/build-agent.ts
+```
+
+## 你应当看到什么
+
+- **第一阶段** —— `default_agent` 在项目下搭建出 `agents/commit-helper/`：它的目录布局、一份复制
+  来的 `system_config.yaml`（含 name + description），以及一份编码了 Conventional Commits 规则的
+  `AGENTS.md`。
+- **第二阶段** —— 这个新的 `commit-helper` Agent，只依据那份生成的 `AGENTS.md`，产出类似这样的结果：
+
+  ```text
+  fix(payment): add retry-with-backoff for transient gateway 503 errors
+
+  Transient 503 responses from the payment gateway were causing checkout
+  failures during peak traffic. Retry with exponential backoff gives the
+  gateway time to recover, preventing spurious user-facing errors.
+  ```
+
+## 说明
+
+- 输出质量取决于模型。`qwen3.6:35b` 能很好地完成这个任务；更小的模型可能无法可靠地遵循工具调用协议。
+- 有能力的模型偶尔会在重新序列化 base config 时引入一个机械性的小失误（例如非法的 YAML 转义）——
+  因为每个文件都是纯文本、每次运行都被追踪，这类失误很容易被发现并修复。这正是“用 Agent 造 Agent”
+  真实的样子：把需求变成 `AGENTS.md` 的重活被自动化了，而机械层面的毛刺由人类来审阅。
+- 再次运行第一阶段会就地更新已存在的 `commit-helper` Agent。
@@ -0,0 +1,82 @@
+/**
+ * Example: an Agent that builds another Agent — programmatically, via the SDK.
+ *
+ * This is the "Harness for Building Agents" pillar in code. It runs entirely on a local
+ * open-weight model (Ollama serving qwen3.6:35b) — see README.md for the one-time setup.
+ *
+ * Two phases, both driven purely through the SDK:
+ *   1. BUILD  — drive `default_agent` with the `agent-creation` skill to scaffold a brand-new
+ *               agent (`commit-helper`) from a plain-language requirement.
+ *   2. RUN    — load the freshly-created agent and have it do its job (write a commit message),
+ *               proving the generated AGENTS.md actually shapes its behavior.
+ *
+ * Run:  pnpm --dir examples/build-agent-with-agent start
+ *   or: npx tsx examples/build-agent-with-agent/build-agent.ts
+ */
+import {
+  createAgent,
+  userText,
+  isCompleteModelMessage,
+  type OmniMessage,
+} from "@prismshadow/penguin-core";
+
+/** Stream a Session run to stdout and collect the assistant's final text. */
+async function runToStdout(run: AsyncGenerator<OmniMessage>): Promise<string> {
+  let finalText = "";
+  for await (const msg of run) {
+    if (isCompleteModelMessage(msg) && msg.payload.type === "text") {
+      finalText += msg.payload.text;
+      process.stdout.write(msg.payload.text + "\n");
+    }
+  }
+  return finalText;
+}
+
+const BUILD_REQUEST = `Use the agent-creation skill to create a brand-new agent in this project.
+
+Requirement: an agent called "commit-helper" that writes high-quality git commit messages.
+Given a diff or a description of changes, it must produce a Conventional Commits message: a
+\`type(scope): subject\` header (type one of feat/fix/docs/refactor/test/chore), subject in
+imperative mood and under 50 characters, then a blank line and a short body explaining the "why".
+
+Do everything the agent-creation skill specifies: create the agent directory layout, copy the
+base system_config.yaml, write a concise AGENTS.md capturing this requirement, and set the new
+agent's name and description in system_config.yaml. Report the files you created when done.`;
+
+const COMMIT_TASK = `Write a commit message for this change: I added retry-with-backoff logic to
+the payment API client because transient 503s from the gateway were causing checkout failures.
+Touched packages/core/src/payment/client.ts.`;
+
+async function main(): Promise<void> {
+  // --- Phase 1: an Agent builds an Agent ------------------------------------------------
+  console.log(
+    "=== Phase 1: default_agent is building a new agent via the agent-creation skill ===\n",
+  );
+  const builder = await createAgent({ agentId: "default_agent" });
+  const buildSession = await builder.createSession({ workspaceDir: process.cwd() });
+
+  // approve: () => "allow" lets the builder run its shell tool calls unattended. In a real
+  // integration you would inspect each tool_call and decide — that is the whole point of the
+  // per-call approval callback.
+  await runToStdout(buildSession.run([userText(BUILD_REQUEST)], { approve: async () => "allow" }));
+
+  // --- Phase 2: run the agent that was just built ---------------------------------------
+  console.log("\n=== Phase 2: running the freshly-created commit-helper agent ===\n");
+  const helper = await createAgent({ agentId: "commit-helper" });
+  const helperSession = await helper.createSession({ workspaceDir: process.cwd() });
+
+  const commitMessage = await runToStdout(
+    helperSession.run([userText(COMMIT_TASK)], { approve: async () => "allow" }),
+  );
+
+  console.log("\n=== Done. commit-helper produced the message above. ===");
+  if (!commitMessage.trim()) {
+    console.error("(No text output — check that Ollama is running and the model is configured.)");
+    process.exitCode = 1;
+  }
+}
+
+main().catch((err) => {
+  console.error("Example failed:", err);
+  process.exitCode = 1;
+});
@@ -0,0 +1,14 @@
+{
+  "name": "@prismshadow/example-build-agent-with-agent",
+  "private": true,
+  "type": "module",
+  "scripts": {
+    "start": "tsx build-agent.ts"
+  },
+  "dependencies": {
+    "@prismshadow/penguin-core": "workspace:*"
+  },
+  "devDependencies": {
+    "tsx": "^4.20.0"
+  }
+}