Claude Code v2.1.198: Background Agents Auto-Submit PRs, 24/7 Unattended Coding

📖 Glossary

AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.

Summary: Anthropic has officially released Claude Code v2.1.198, with the core upgrade being that subagents now run in the background by default — while you discuss a plan with Claude, it automatically completes code refactoring, test runs, and PR submissions in the background. Combined with the previously launched Routines (cloud-based scheduled tasks) and Dynamic Workflows (dynamic orchestration of dozens to hundreds of parallel subagents), Claude Code is evolving from "a programming tool you actively use" into "a coding agent that works for you 24/7."

Three Capability Pieces Complete the Puzzle

To understand the significance of v2.1.198, you need to connect Claude Code's three key updates over the past three months.

Kaihe AIBOX has been closely watching the autonomous execution capabilities of AI Agents, and this Claude Code upgrade marks a milestone in Agent autonomy.

The first piece is Routines, launched on April 14. A Routine equals a prompt plus a codebase plus connectors (GitHub, Slack, Linear). Configure it once, and it runs automatically on Anthropic's cloud — no need for your computer to be on. Three trigger types: scheduled (pull bugs and open Draft PRs every night at 2 AM), API calls (trigger smoke tests after CI/CD completes), and GitHub Webhooks (auto-trigger Code Review on new PRs). Pro users get 5 runs/day; Team/Enterprise gets 25.

The second piece is Dynamic Workflows. Claude automatically writes JavaScript orchestration scripts, dispatching dozens to hundreds of parallel subagents for large-scale tasks — library migrations across hundreds of thousands of lines of code, full security audits, multi-angle technical validation. Each subagent has its own 200K token context window and independent tool permissions, with intermediate results staying out of the main conversation.

The third piece is v2.1.198: subagents run in the background by default. Previously, you had to manually press Ctrl+B to move a task to the background. Now it's reversed — background is the default; foreground requires you to explicitly say so. You discuss a refactoring plan with Claude while it simultaneously runs tests and submits PRs in the background.

What v2.1.198 Actually Changes

Many might think "default background" is just a UI tweak. But anyone who's actually used it knows this is a fundamental shift in operating mode.

The old workflow was serial: ask Claude to run tests → wait → review results → decide next step → ask Claude to modify code → wait again. One feature from start to PR submission might require five or six rounds, each taking minutes. The human was always waiting for the machine.

Now it's parallel: you discuss "how to refactor this module better" with Claude, and while you're conversing, a background subagent is already running existing tests, analyzing dependencies, and generating a refactoring PR. By the time your discussion concludes, the PR might already be open. The machine works while waiting for the human — and while the human thinks.

This is what Anthropic means by the "orchestrator" position — you're no longer the executor, or even the waiter. You're the orchestrator managing multiple parallel task streams.

How It Differs from Routines

You might ask: isn't Routines also automatic execution? What's different about v2.1.198's background agents?

The difference lies in trigger mechanisms and autonomy level.

Routines are preset and structured — you define the prompt and trigger conditions in advance (schedule, API, Webhook), and it executes by rules. Suitable for repetitive, predictable tasks. Like pulling bug lists every night and fixing them, or running smoke tests after each deployment.

v2.1.198's background agents are immediate and adaptive — you mention "help me refactor this module" in conversation, and Claude decides what subtasks to create, which can run in the background, and which need your confirmation. It doesn't execute a preset script; it understands your intent in real-time and schedules autonomously.

The official store has hardware solutions suited for running these types of Agent workloads.

Real-World Scenarios

Scenario 1: Large-scale library migration. You have 500K lines of code to migrate from Vue 2 to Vue 3. Previously, this required two to three weeks of manual engineering. With Dynamic Workflows, Claude generates an orchestration script, dispatches hundreds of parallel subagents — each handling one module's migration: analyzing component dependencies, generating migration code, running tests, aggregating results. You wake up, and Draft PRs are ready.

Scenario 2: Unattended bug fixing. Configure a Routine that pulls top-priority bugs from Linear every night at 2 AM. Claude analyzes each bug, locates the code, generates a fix, opens a PR, and pushes a summary to Slack. You arrive at work in the morning, and the PR list is waiting for Review. Pro's 5 runs/day is enough for one such task.

Scenario 3: Real-time collaborative refactoring. You're discussing a core module refactor with Claude. Halfway through, you want to check existing test coverage. Previously, you'd have to stop and wait. Now you say "also run the existing test coverage," a background subagent starts immediately, you continue discussing the plan, and results come back in a few minutes — just as the discussion wraps up.

AI coding agents working in parallel

Usage Limits and Cost

Current Routines daily limits: Pro 5/day, Max 15/day, Team/Enterprise 25/day. Background subagent token consumption is billed like normal conversation, but since it runs in an independent context window, it doesn't squeeze the main conversation's token budget — it's actually more efficient, since intermediate results don't need to return to the main conversation.

Dynamic Workflows has no hard cap on parallel subagents, but each is an independent API call. Hundreds of parallel agents means hundreds of API calls' worth of cost. For large-scale tasks, start small and scale up.

What This Means for Developers

Simply put: programming is shifting from "you sit at a computer writing code line by line" to "you design tasks, review results, and manage multiple parallel streams."

The core competency is shifting. It used to be coding speed and quality. Now it's increasingly about: can you decompose a complex task into subtasks that Agents can execute, can you design good prompts and orchestration logic, can you quickly Review and judge Agent output quality.

It's not that coding doesn't matter anymore — it's that "writing code" itself is being taken over by Agents, and human value is migrating upstream (task design and quality control) and downstream (system integration and business logic).

Unattended PR submission notification

Risks and Limitations

Background Agent code quality isn't guaranteed. Claude Code's agents still make mistakes on complex business logic, especially in domains requiring deep domain knowledge. The PR Review step cannot be skipped — at least not yet.

When Dynamic Workflows dispatches hundreds of subagents, poorly designed orchestration scripts can produce massive duplicate work or even conflicting modifications. You need a clear understanding of the task structure — you can't just hand it to Claude and walk away.

Routines is still in Research Preview, with tight limits and iterating features. Pro's 5 runs/day is indeed limited — best to start with one high-frequency scenario.

Claude Code v2.1.198: Background Agents Auto-Submit PRs, 24/7 Unattended Coding