Anthropic releases ant CLI for Claude Platform and Managed Agents tracing

• Anthropic shipped ant, a terminal client that exposes the full Claude API as subcommands, with typed flags, piped YAML, @path file references, and built-in response transforms, according to the CLI docs and ClaudeDevs' launch post.
• Managed Agents support is part of the same terminal flow: ClaudeDevs' tracing demo shows a session kicked off from the CLI, while Anthropic's events docs describe retrievable session, span, and agent events with server-side timestamps.
• Claude Code can already drive ant through its built-in /claude-api skill, so the coding agent can list sessions, upload PDFs, and debug runs by shelling out to the same CLI, per ClaudeDevs' /claude-api post.
• Anthropic also changed Claude Code's session handoff: ClaudeDevs' /fork update says /fork now launches a background agent with the same system prompt, tools, history, model, and prompt cache, while /branch inherits the old transcript-copy behavior.
• The background-agent push lands a day after Anthropic reset Claude Code limits for Pro and Max users because Opus 4.8 was spawning too many parallel tool calls, as ClaudeDevs' rate-limit reset post and ClaudeDevs' follow-up explained.

You can browse the CLI docs, inspect the Managed Agents event model, and read Anthropic's new dynamic workflows post. There is also a public anthropic-cli repo and release feed, which shows the tool has been shipping steadily for weeks before this bigger push into Claude Code.

Anthropic's pitch is straightforward: the platform has a terminal now. The ant docs say every API resource is exposed as a subcommand, so the same binary can hit Messages, files, sessions, and other endpoints from shell scripts or interactive runs.

The mechanics are more opinionated than a thin curl wrapper. According to the CLI docs, ant lets you build requests from typed flags or piped YAML, inline file contents with @path, auto-paginate list endpoints, and pull fields back out with --transform.

Install options are already spread across the usual developer surfaces. ClaudeDevs' install post points to brew, curl, and Go, while the public GitHub releases page shows signed binaries and versioned changelogs.

The more interesting part is not basic API access, it is observability. Anthropic's demo shows a CLI flow where you start a Managed Agents session, send it an event, wait for it to work, then pull the trace back from the same terminal.

The docs match that framing. The sessions reference says sessions can pin an agent version or mount resources like repositories, files, and memory stores, and the events docs describe three visible layers:

• user events, which kick off or steer a run
• session, span, and agent events, which expose status and progress
• processed_at timestamps, which show whether an event has been recorded server-side or is still queued

That makes ant look less like a convenience wrapper and more like a debugging surface for long-running agent runs.

Anthropic's companion blog post on dynamic workflows fills in why the company is suddenly talking so much about background orchestration. The post says Claude Code can now write its own task-specific harness when the default single-context loop starts to break down on long-running, highly parallel, structured, or adversarial work.

The patterns Anthropic highlights are unusually explicit. The blog describes workflows as a way to counter what it calls agentic laziness, and ClaudeDevs' demo shows the practical version: encode your manual checks so Claude closes its own feedback loop before returning an answer.

One detail from [trq212's post](src:4|trq212's workflows post) is worth keeping: he called workflows the biggest Claude Code capability jump since skills and subagents, and said he was especially interested in the non-technical tasks they unlock. The official post does not spend much time on that angle, but it fits the broader pattern of Anthropic moving orchestration features out of pure coding workflows and into reusable task harnesses.

Anthropic changed one of Claude Code's old verbs at the same time. /fork now means, "run this in the background with my current setup," not, "copy this transcript into a new session."

According to ClaudeDevs' command update, the new /fork inherits five pieces of state:

• system prompt
• tools
• conversation history
• model
• prompt cache

/branch takes over the old behavior, which is a plain transcript copy into a new session you keep driving yourself. That is a small naming tweak with a real workflow consequence: Anthropic is turning session handoff into delegated execution, not just session management.

Anthropic also spent part of this launch window cleaning up a failure mode from the previous week. The company said it reset 5-hour and weekly Claude Code limits for Pro and Max users after some sessions spawned excessive parallel subagents and burned usage faster than expected.

The follow-up matters because it narrows the blame. ClaudeDevs' clarification said the issue was tied to how Opus 4.8 requests were handled, which caused more parallel tool calls than intended, and said it was unrelated to dynamic workflows.

That leaves an awkward but useful read on the week's releases: Anthropic is shipping more ways to fan work out in the background, while also having to patch the cost of too much background work one day earlier.