Safety controls and SDK hooks
Anthropic added strict execution controls to Claude Code v2.1.317. When integrating with the Agent SDK, developers can specify an onFailure setting of block for hook callbacks. This configuration ensures that any failed hook callback actively halts the agent action rather than allowing execution to continue silently.
The release also introduces an interactive prompt before web or shell tool calls send an account email address across the network. In addition, auto mode now alerts users when a subagent tool invocation produces no answer, replacing earlier behavior that left the primary view empty.
Runner isolation and subagent coordination
Anthropic patched nested session collisions on runners where executing sub-commands shared session IDs, wrote into transcripts, and triggered erroneous conversation resumes. Teams wanting earlier behavior can enable the CLAUDE_CODE_DISABLE_NESTED_SESSION_ID environment variable on self-hosted runners.
Multi-agent coordination received several behavioral fixes. Teammates running in tmux or iTerm2 panes no longer launch in auto mode when the lead session operates in manual mode. Meanwhile, incoming messages to background agents now move running processes to the background instead of halting them, and subagents on different models default to their own saved effort levels.
Our assessment
The addition of blocking hooks provides a reliable way to enforce external authorization policies before tools execute. Engineering teams building around the Agent SDK should verify whether existing hooks rely on silent failure fallbacks. Operators running self-hosted runners should audit nested session scripts to confirm whether isolating transcripts changes their debugging workflows.
Source: Anthropic original publication.
Read the original material
Anthropic (opens in a new tab)The article date belongs to Agentic Horizon. The source publication date is listed separately. Vendor claims remain attributed; we have not independently tested the reported capability.
