Claude Code: Agent View, Goal Completion, and a Hardening Governance Surface

What Changed

Claude Code shipped three interrelated surface expansions in v2.1.133--v2.1.139 (May 7--11, 2026).

Agent view (claude agents, Research Preview, v2.1.139) is a full-screen supervisor panel for all Claude Code sessions. Sessions are organized by state -- working, needs input, completed, failed, stopped -- under a persistent supervisor process that keeps sessions alive when no terminal is attached. File edits in background sessions are automatically isolated to per-session git worktrees under .claude/worktrees/. A peek panel lets operators reply without full attach; the peek panel shows what the session needs and accepts multiple-choice responses by key. Session rows display Haiku-generated one-line summaries, refreshed at most every 15 seconds while the session is active. The disableAgentView managed setting gives enterprises the option to disable the surface. Background sessions survive terminal closure; claude respawn --all restarts stopped sessions after machine sleep.

/goal command (v2.1.139): sets a completion condition; Claude continues working across turns until it is met. Works in interactive, -p, and Remote Control modes. A live overlay shows elapsed time, turns, and token consumption. This is the first explicit goal-completion primitive in Claude Code.

Hook continueOnBlock for PostToolUse (v2.1.139): when set to true, feeds the hook's rejection reason back to Claude and continues the turn rather than ending it. Governance hooks can be advisory -- a block is a constraint Claude can respond to -- rather than terminal.

Subagent attribution headers (v2.1.139): all API requests from subagents carry x-claude-code-agent-id and x-claude-code-parent-agent-id HTTP headers. OTel claude_code.llm_request spans include agent_id and parent_agent_id attributes. Multi-agent call trees are now attributable at the API layer.

API key disables cloud surfaces (v2.1.139): when ANTHROPIC_API_KEY, apiKeyHelper, or ANTHROPIC_AUTH_TOKEN is set, Remote Control, /schedule, claude.ai MCP connectors, and notification preferences are disabled even if a Claude.ai login also exists. Cloud control surfaces require cloud auth.

settings.autoMode.hard_deny (v2.1.136): auto mode classifier rules that block unconditionally regardless of allow exceptions. Hard blocks that no allow rule can override.

worktree.baseRef (v2.1.133): controls whether --worktree, EnterWorktree, and agent-isolation worktrees branch from origin/<default> (fresh, the new default) or local HEAD (head). This reverses the default that was in place since v2.1.128.

Hooks receive effort level (v2.1.133): all hook handlers receive the active effort level via the effort.level JSON input field and the $CLAUDE_EFFORT environment variable.

Operator Consequence

Claude Code now has the scaffolding for a supervised multi-agent operating style. Agent view makes session state an observable surface rather than ambient assumption. /goal lets operators hand off a completion condition without babysitting each turn. Together they address the two main barriers to multi-agent adoption: not knowing what all sessions are doing, and not knowing when to stop.

The governance additions complete a policy model for auto mode: hard_deny + soft deny + allow rules. continueOnBlock closes the gap between "governance stops the agent" and "governance teaches the agent." The API key boundary makes an implicit behavior explicit, reducing the chance of discovering missing features at runtime.

Subagent attribution headers are a small change with compounding value: any logging pipeline that receives Anthropic API calls can now distinguish parent-session requests from subagent requests and reconstruct call trees.

The worktree.baseRef: "fresh" default change is a breaking behavior change for operators who relied on unpushed commits being present in new worktrees. Set worktree.baseRef: "head" to restore the prior behavior.

Bitter Implication

Bitter adapters for Claude Code should capture:

• Session state from the supervisor roster (~/.claude/daemon/roster.json) when wrapping multi-session runs
• x-claude-code-agent-id and x-claude-code-parent-agent-id headers to reconstruct agent call trees in BitterLog
• Goal text and termination reason when /goal is in use, as these constitute operator intent evidence
• Hook rejection reasons when continueOnBlock is active -- these form a governance event stream that belongs in the run receipt

The capability-profile assumption for Claude Code worktree behavior should be updated: the default base is now the remote default branch, not local HEAD.

Signal

Agent view, /goal, and continueOnBlock together represent a structural expansion: Claude Code is building the surfaces an operator needs to run it unsupervised at scale. This changes what Bitter should test (supervisor-based session wrapping), what capability profiles should assert (goal completion is native), and how governance hooks should be designed (advisory rejection with reason is now possible). Accessible enough to change operator behavior now; material enough to change Bitter adapter assumptions.