Skip to content

The Agent panel

View → Agent (Ctrl+Alt+I) opens a chat to the right of the editors. It is an Agent Client Protocol client — the protocol Zed's agent panel speaks — not a model call: Arvo launches an agent as a subprocess, hands it Arvo's research tools, streams its turn into the panel, puts its permission requests to you, and serves its file reads and writes inside the project folder.

By default the agent is Claude Code, through the protocol's adapter (npx @agentclientprotocol/claude-agent-acp), using your own Claude login. The agent.command setting names any other ACP agent.

What the agent can do

  • read findings, list rules and rulesets, write a ruleset, run a study or a walk-forward, look at the bars a study saw and the regime each closed in, and read findings against each other — Arvo's MCP tools, behind the research token, which has no call that fetches data, touches a credential or trades;
  • read and write files in the project (a ruleset it writes appears in the editor);
  • run commands in the project folder, each behind a permission prompt, with the output under the tool card.

What it cannot do, by construction: reach an account, an order, or anything outside the project.

Where the agent fits in automated trading

Arvo's end goal is automated trading: a rule that survived research runs against a venue with no person at each order. That is what a session is. The agent's role is to set up the rules that trade — author rulesets, run studies, read the advice, and in time promote a finding to a session and adjust it from the daily reviews. The rule trades; the agent researches and operates. Every order is still one the gate sized and the record can explain back to a bar and a rule.

The other meaning — the agent looks at the market and places an order because it decided to — is deliberately not built. An order like that has no rule behind it to backtest, no finding to go stale and no search to deflate against; nothing explains it but a transcript. If it is ever built, it goes through the gate like any other proposer, and it takes a decision that supersedes the current one. Where this is going →

What the panel shows

The agent's words; each tool call as a card with its state, the files it touched as links into the editor and the finding it produced as a link into Runs; permission prompts with the agent's own options (allow once, allow always, reject). A finding the agent makes appears under agent chat in the Runs tab beside the scripts' groups.