Anthropic adds `ant apply` to ant CLI, letting developers manage Claude Managed Agent resources as repo files

Social·Community

Claude Managed Agents Workshop Q&A: CI/CD, Tenancy, GPUs and Coordinator Pattern (Parts 3–4) — Anthropic

Anthropic engineers Priyanka Phatak and Gabe Cemaj ran a four-part Claude Managed Agents workshop at AI Engineer World's Fair 2026, with parts 3 and 4 posted on October 12. Part 3 demos scheduled deployments, multi-session fan-out, and Dreaming, a research preview that reviews 30 days of past sessions to improve a memory store. Part 4 covers versioning agent configs in YAML via CI/CD with the ant CLI, tenant isolation, self-hosted GPU sandboxes, and a coordinator pattern for custom tools.

Anthropic engineers Priyanka Phatak and Gabe Cemaj ran a four-part Claude Managed Agents workshop at AI Engineer World's Fair 2026, with parts 3 and 4 posted on October 12. Part 3 demos scheduled deployments, multi-session fan-out, and Dreaming, a research preview that reviews 30 days of past sessions to improve a memory store. Part 4 covers versioning agent configs in YAML via CI/CD with the ant CLI, tenant isolation, self-hosted GPU sandboxes, and a coordinator pattern for custom tools.

Releases·Tool Release

Anthropic Decouples Claude Managed Agents' Harness from Sandboxes, Cutting Time to First Token ~90%

A workshop video from AI Engineer World's Fair 2026 shows Priyanka Phatak, who leads the Claude Managed Agents engineering team at Anthropic, walking through the platform's architecture redesign. The first version ran session, harness, and sandbox in one hard-to-debug 'pet' container; the redesign decouples a brain (harness plus Claude) from hands (sandboxes and tools) around an append-only session log. Practitioners get a documented pattern for keeping harness assumptions from going stale with each new model.

A workshop video from AI Engineer World's Fair 2026 shows Priyanka Phatak, who leads the Claude Managed Agents engineering team at Anthropic, walking through the platform's architecture redesign. The first version ran session, harness, and sandbox in one hard-to-debug 'pet' container; the redesign decouples a brain (harness plus Claude) from hands (sandboxes and tools) around an append-only session log. Practitioners get a documented pattern for keeping harness assumptions from going stale with each new model.

A guide circulating on X walks through deploying a Claude Managed Agent that reads Slack and GitHub on a schedule, with credentials and memory, and posts updates. Setup runs through a Claude Code command. It also notes that the new API credits are included in Max and Team plans.

Social·Community

Claude Managed Agents completed status can still include failed work

Anthropic's Managed Agents workflow-run documentation shows a run can return a completed status even when work on a thread failed or a thread could not be created, with failures only visible in individual thread events. This matters for practitioners building dynamic workflows, since a completed run may not mean full task coverage. The post proposes surfacing coverage and acceptance checks separately from execution status.

Anthropic's Managed Agents workflow-run documentation shows a run can return a completed status even when work on a thread failed or a thread could not be created, with failures only visible in individual thread events. This matters for practitioners building dynamic workflows, since a completed run may not mean full task coverage. The post proposes surfacing coverage and acceptance checks separately from execution status.

Agentctl is a tool for managing configuration of multiple CLI coding agent tools (opencode, claude, agy) declaratively, like Terraform. It lets users keep consistent setups across different agent harnesses and sync that configuration between machines. It targets developers who use several agent CLIs and don't want to configure each separately.