Human deciders and tool permission boundaries
Anthropic added a decider setting to mod tool evaluations, allowing developers to enforce that only a direct permission dialog or an authorized host callback can validate a tool request. If an automated environment runs unattended with nobody present to review the prompt, the agent runtime automatically rejects the invocation using the explanation provided by the mod. The tool evaluation event also now exposes whether an attempted action is read-only.
The update also caps skill hook scale, restricting any single repository, plugin, or user folder to registering 1,024 hooks and 16,384 matcher characters in a given session. Additionally, mod configuration hooks can inspect each layer of incoming configuration to rewrite combined values, enabling enterprise administrators to impose hard ceilings on values like retention sweep periods.
Subagent lifecycles and broader deployment targets
Claude Code v2.1.304 addresses edge cases where plugins kept headless operations running past configured step caps, stop commands, or finished inputs by instantiating subagents or re-executing finished ones. It also resolves a bug in multi-agent team setups where confirming an agent shutdown from the primary session accidentally killed the entire command-line process.
The release expands environment adaptation with a dedicated command that inspects local workspaces and historical sessions to calibrate auto mode. Anthropic made this guided setup accessible across deployment routes including Amazon Bedrock, Google Vertex AI, Microsoft Foundry, LLM gateways, and environments operating without telemetry.
Our assessment
Teams managing automated runs should evaluate whether unattended pipelines break on mod tool checks that now require human confirmation. If background workflows rely on plugins spawning subordinate workers to finish jobs, operators should verify that workflows conclude within standard step limits rather than running past normal termination bounds.
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.
