The chatbot (module 25) answers from the model's head: one completion call. This module gives the agent hands. Its @Prompt handler hands the user's goal to an agent loop that reads and edits files and runs tools in the open project — for as many turns as the goal needs — then reports back.

The ACP skeleton is identical to the chatbot (@Initialize / @NewSession / @Prompt, same as module 25). The only change is the body of @Prompt:

AgentClient.create(model)
    .goal(request.text())
    .workingDirectory(cwd)
    .run();

That's agent-client — the AgentClient abstraction. agent-claude drives the local Claude CLI as the underlying model.

echo (12) → completion chatbot (25/26/27) → AGENTIC (this, 32) → bud (final boss)

The chatbot calls the model once. bud is a full domain buddy (knowledge base, curated review workflow, provenanced report). This module is the honest middle step: a real agent loop wired into one ACP handler — and nothing more.

This is not bud. No knowledge base, no curated report, no multi-step named workflow, no judge. Those are what make bud bud. This is just agent-client in a @Prompt method — the smallest example of "an ACP agent that is an agent, not a chatbot."

The line to land: "The chatbot answers from the model's head. This one has hands — it reads your files and edits them. Same three ACP handlers; that's the jump from chatbot to agent."

• The Claude CLI (Claude Code) installed and logged in. agent-claude drives that CLI — it does not call the Anthropic HTTP API directly. Run claude once to confirm it opens without asking for a key.
• Billing / no API key: the CLI uses your Claude subscription when ANTHROPIC_API_KEY is absent, and the API when it's present. So — unlike the chatbot — run this agent without ANTHROPIC_API_KEY in its environment. In the IDE, launch it via the acp-no-key.sh wrapper (see below), not the key-injecting one.
• Java 17.

./mvnw package -pl module-32-agent-client -q
./mvnw exec:java -pl module-32-agent-client

The demo launches the agent in a throwaway working directory and gives it an agentic goal — "create a haiku.txt file …" — then reads the file back from disk to prove the agent actually wrote it (a chatbot could only describe one).

Plugged into IntelliJ like the other agents (see module 29), but with the no-key wrapper so the underlying claude CLI uses your subscription:

"Agent-Client": {
  "command": "/path/to/acp-no-key.sh",
  "args": ["java", "-jar", "/abs/path/to/module-32-agent-client/target/agent-client-agent.jar"]
}

where acp-no-key.sh simply unsets ANTHROPIC_API_KEY and execs its arguments. It reviews / edits whatever project the IDE window has open, and — like bud — the run is agentic and slow (multiple tool-using turns), not instant like the chatbots.

agent-client isn't Claude-only: agent-claude has siblings (agent-gemini, agent-codex, …). Build a different AgentModel and the rest of the handler is unchanged — the agent loop, and the ACP wire protocol, stay the same.