Anthropic Turned Claude Code Into an Unattended Agent Runtime This Spring

When Anthropic released Claude Opus 4.8 on May 28, it bundled in a Claude Code feature called dynamic workflows that can run hundreds of subagents in a single session and, in the company's own example, carry a code migration across hundreds of thousands of lines "from kickoff to merge." The launch capped a spring in which Claude Code gained the ability to run for long stretches without pausing to approve each edit.

Across versions 2.1.72 to 2.1.166, published between March 9 and June 5, Anthropic converted a terminal program into a multi-surface agent runtime built to run without a human watching. The harder engineering, and the part that matters most for any company deciding whether to let an agent into its repositories, was the policy layer shipped alongside the autonomy, the classifier gates, hard-deny rules and managed settings that decide what an unattended agent is allowed to touch.

Auto mode and the classifier it runs on

Auto mode arrived in late March as a research-preview permission mode, a middle setting between approving every action and the blanket --dangerously-skip-permissions flag, according to Claude Code's Week 13 release notes. The current permission-modes documentation is more precise. Safe reads and working-directory edits run, while a separate classifier model reviews anything beyond that, blocks actions that exceed the request or target unrecognized infrastructure, and sends denial reasons back so Claude can try another path.

Anthropic's documentation does not name that classifier, describing it only as server-configured and independent of the model running the session. For hard guarantees, the configuration page tells administrators not to lean on auto mode's trust settings at all, directing them instead to a permissions.deny rule in managed settings, which the docs say "blocks the action before the classifier is consulted and cannot be overridden." The trust configuration is deliberately asymmetric. The classifier reads settings from user, local and managed scopes but not from a repository's shared settings file, so a project cannot grant itself broad permissions by committing a config.

Background work got a supervisor process

In version 2.1.139, Claude Code added an agent view, a dashboard reached with the command claude agents that lists background sessions as working, waiting, idle, completed or failed. The hosting model changed underneath it. A per-user supervisor process now starts when a session is backgrounded, each background session runs as its own Claude Code process that the supervisor keeps alive without the launching terminal, and state persists in the Claude Code config directory, by default under ~/.claude/jobs. A companion command, /goal, sets one completion condition that a small, fast model checks after every turn, ending the per-turn prompt the way auto mode ended the per-tool prompt.

Anthropic's Opus 4.8 launch post led with what it called honesty, claiming the model is "around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked." Scott Wu, chief executive of Devin maker Cognition, said in a testimonial Anthropic published that the model "follows instructions with the consistency our autonomous engineering workloads need to keep running unattended."

Four execution surfaces

Work can now leave the terminal in four ways, and the docs give each its own access rules. Sessions run locally on the command line, on Anthropic-managed cloud machines at claude.ai/code, through a Remote Control mode that keeps execution local while a browser or phone acts as the screen, or as cloud-run routines, plan drafting and code review. Anthropic frames all of it as capability, calling Opus 4.8 a "more effective collaborator" and workflows a way to take on "even bigger tasks."

Auto mode, workflows and every cloud surface carry a research-preview label. Team and Enterprise administrators must switch on auto mode and Remote Control before anyone can use them, organizations with zero-data-retention contracts are shut out of cloud sessions and remote review, and the Bedrock, Vertex and Foundry deployments many regulated buyers run cannot use fast mode at all. Fast mode itself got cheaper, falling to $10 and $50 per million input and output tokens on Opus 4.8 from $30 and $150 on the previous models, while regular pricing held at $5 and $25. Even so, Claude Code had reportedly taken up to 54% of the AI coding market by April, per Menlo Ventures data The Implicator cited at the time.

Most of the features that let Claude Code run unattended are still labeled research preview, and the strongest controls live in managed settings only an administrator can set. Anthropic's June changelog shows where the work is heading. Between versions 2.1.160 and 2.1.166 the company renamed the workflow trigger to "ultracode," added managed minimum and maximum version requirements, and hardened cross-session messaging so that a backgrounded agent now refuses relayed permission requests. The release that turns these controls on by default is the one that will tell enterprises whether to let an agent work in their repositories while no one is watching.

Professional Tutorial: Build a Pi-to-Pi Agent Communication Bus

The useful version of multi-agent coding does not start with ten agents. It starts with two Pi Coding Agent sessions, using the terminal coding harness to hold different context: one session sits near

The Implicator

Repo Radar: 5 GitHub Projects Worth Your Week

Repo Radar's fourth issue leaves the coding agents alone and looks at the tooling around them. The five projects below were all pushed within the last 48 hours. Memory between sessions, code indexing,

The Implicator

Pi Is Not a Claude Code Rival. It Is a Harness Rebellion

Flask creator Armin Ronacher published a technical essay on his personal site on January 31 that endorsed a yellow terminal coding agent called Pi as "the minimal agent within OpenClaw." His company E

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai