Skip to content
Playground

Workspace-Scoped Coding Agent

Last updated today · 51d57ea

↑ Updated today

"docs: v0.18.0 release documentation" · 51d57ea · 2026-11-07

  • ## Scope the agent to a project
  • ## Switch projects per run
  • ## Load a skill mid-run
  • ## Durable version

A coding agent needs three things: a root it cannot escape, the project’s own instructions, and limits on the tools that touch the outside world. In v0.18 these are all per-run, so one agent instance can work on several projects without leaking state between them.

import { ReactiveAgents } from "reactive-agents";
const agent = await ReactiveAgents.create()
.withProvider("anthropic")
.withTools({
builtins: ["file-read", "file-edit", "grep", "list-directory"],
workspace: "/home/me/projects/billing-api", // must be an existing directory
maxCallsPerTool: { "file-edit": 20, grep: 40 },
riskLevels: { grep: "low" }, // lets the text parser bind embedded grep calls
})
.withSkills() // no paths: discovers the workspace's skills
.build();
const result = await agent.run("Find where invoices are rounded and fix the off-by-one");
console.log(result.output);
await agent.dispose();

What this gives you:

  • File tools stay inside workspace. Relative paths resolve against it, and paths outside it are refused.
  • Skills come from the project. .withSkills() with no arguments loads <workspace>/.agents/skills, <workspace>/.<agentId>/skills, and <workspace>/.claude/skills. The agent sees the project’s own conventions. User-global skills (~/.agents/skills) load only with .withSkills({ global: true }).
  • Caps bound the damage. file-edit runs at most 20 times in the run, however the model splits its calls across turns.

The workspace is a run option. Override it without building a second agent:

await agent.run("Add retry logic to the webhook sender", {
workspace: "/home/me/projects/notifications",
});

Precedence is: the run option, then .withTools({ workspace }), then an ambient withFileRoot(). The skills for that run are discovered from the run’s workspace, so the second project’s .agents/skills apply to that run only.

Skills in the catalog are not all injected up front. The agent can call activate-skill to load one when the task needs it. A skill loaded this way is kept through context compaction, so it does not fall out of the conversation on long runs.

Add .withDurableRuns() to make the run resumable. Any file edit that succeeded before a crash is recorded as a tool effect and is not re-applied on resume. See Durable Execution.