Claude Code 2.1.178-185: Auto Mode Safety and Agent Teams Overhaul
Anthropic shipped five Claude Code releases between June 15 and June 20, delivering a substantial set of safety improvements, configuration enhancements, and stability fixes. The most visible change arrives in v2.1.183, which adds safety gates to auto mode: destructive git operations like git reset --hard, git checkout -- ., git clean -fd, and git stash drop are now blocked unless you explicitly asked to discard local work. Similarly, git commit --amend is blocked when the commit was not made by the agent in the current session, and infrastructure teardown commands for Terraform, Pulumi, and CDK are blocked unless you requested the specific stack.
Version 2.1.178 restructures Agent Teams by removing TeamCreate and TeamDelete tools entirely. Every session now has one implicit team, and teammates are spawned directly via the Agent tool's name parameter with no setup step. The same release introduces Tool(param:value) syntax for permission rules, allowing fine-grained control over tool inputs with wildcard support, such as Agent(model:opus) to block Opus-model subagents. Auto mode also gains a classifier that evaluates subagent spawns before launch, closing a gap where subagents could request blocked actions without prior review.
Version 2.1.181 adds /config key=value shorthand syntax for setting any configuration from the prompt, upgrades the bundled Bun runtime to 1.4, and improves streaming behavior so long paragraphs appear line-by-line instead of waiting for the first line break. Connection reliability gets a boost with automatic retries for API drops that occur mid-thinking, replacing the previous error display.
Across all five releases, over 40 bug fixes address issues ranging from prompt caching failures on custom API endpoints, to file corruption on network drives, to fullscreen TUI rendering problems under heavy subagent load on Windows Terminal.
Read more — Claude Code Docs
Hermes v0.17.0: The Reach Release Expands Communication Channels
NousResearch shipped Hermes Agent v0.17.0, dubbed "The Reach Release," on June 19, expanding the personal AI assistant across multiple new communication channels. The release adds native iMessage support via the Photon adapter, eliminating the previous requirement for a dedicated Mac relay, and introduces an official WhatsApp Business Cloud API adapter alongside rich text support for Telegram messages.
The release integrates with Raft agent networks as a gateway channel, enabling Hermes to participate in multi-agent coordination across distributed agent meshes. Background and asynchronous subagent delegation allows long-running tasks to continue without blocking the primary conversation thread. The desktop app receives substantially enhanced watch-windows for monitoring subagent activity in real time.
New Automation Blueprints provide a friendlier alternative to raw cron syntax for scheduling recurring tasks, while the image editing capabilities added to the image_generate tool expand creative workflows. The Skills Hub browser has been redesigned, and the Curator cost optimization eliminates auxiliary model spend on routine maintenance runs. The release closes over 300 issues in total.
Read more — GitHub
Claude Agent SDK Introduces Separate Usage Credits
Starting June 15, 2026, Anthropic began drawing Claude Agent SDK usage and non-interactive claude -p runs from a dedicated monthly Agent SDK credit pool, separate from interactive session limits. Pro subscribers receive $20 in monthly agent credits, while Max 5x and Max 20x plans receive $100 and $200 respectively.
The billing separation reflects the growing divergence between interactive coding sessions and autonomous agent workloads. Interactive use involves real-time human-in-the-loop collaboration where token consumption is bounded by conversation pace, while agent SDK runs can consume significantly more tokens per unit of time through parallel tool execution and background processing. By splitting the credit pools, Anthropic avoids a scenario where heavy agent automation drains a developer's interactive allocation.
For teams running CI/CD integrations, scheduled code reviews, or automated workflows through claude -p, the dedicated credit pool provides predictable capacity that does not compete with day-to-day coding assistance. The change also signals that Anthropic views agent-mode usage as a distinct product surface with its own economics.
Read more — Claude Code Docs
