AI weekly news (2026-05-12): Claude, Codex, Gemini, MCP — what landed

"AI sales automation still needs you to give instructions every turn." That was the limit of AI sales automation as of 2026-05-11. On 2026-05-12, two features removed that wall.

Anthropic released Claude Code v2.1.139, which contains 50 CLI changes. Two of them are pivotal:

1. /goal slash command — "continue execution across turns until a condition is met," natively in the CLI
2. Agent View (claude agents) — monitor and control multiple parallel sessions from one screen

Primary sources: Anthropic official documentation, GitHub Releases, official Claude X. OpenAI official information is referenced for adjacent industry context.

1. The /goal command — "AI that doesn't stop until done"

Traditional Claude Code is conversational. The user instructs, Claude responds. For complex tasks, the user has to approve and add instructions every turn.

The /goal command flips this assumption.Set the condition once, and Claude Code keeps calling tools and reasoning autonomously until the condition is judged "done."

Execution model — two AIs split the work

The implementation is elegant:

1. Judge AI: at end of each turn, the current context + your condition is sent to a small model (default: Haiku). It returns { done: true|false, reason: string }.
2. Worker AI: if done: false, the reason becomes next-turn guidance for the main Claude Code, and the Worker continues.

The community has called this pattern the "Ralph Loop."Karpathy implemented it for autoresearch; Anand Chowdhary's continuous-claude shipped a version. This is the first time Anthropic itself baked it into the CLI.

Three execution modes for /goal

• Interactive mode (regular launch) — type /goal in chat
• -p Pipe mode — one-shot with claude -p "/goal ..."
• Remote Control mode — external tools (like Sales Claw) can inject goals via API (the key for Sales Claw integration)

Basic usage

Launch Sales Claw, then enter:

# Inside Claude Code's interactive mode
/goal Process the target list from the top until 10 companies are awaiting approval

# Or one-shot via -p (pipe)
claude -p "/goal Process until awaiting_approval reaches 10 companies. Skip do-not-contact; register CAPTCHA into awaiting_approval"

That's all. Claude Code loops form-fill → screenshot → log → next company on its own until 10 awaiting_approval entries accumulate.

Example — automating prospect list creation

Feed this /goal to research and qualify a list of SaaS companies in Tokyo:

/goal List 50 SaaS companies in Tokyo with 50–500 employees.
For each company, fetch the contact-form URL, judge CAPTCHA presence,
check for "no sales" notices and assess contact viability, save to results.csv

Claude Code completes the 50-company research, classification, and CSV writeout in a single /goalrun. The live panel shows "N remaining" and "currently company X."

Best practice + caveat

• Judge model: Haiku 4.5 recommended (fast, low cost, accurate enough)
• For long loops, also use --max-turns
• Always stop send-bearing operations at awaiting_approval and log to the audit trail

2. Agent View

What Agent View is

Running multiple Claude Code sessions in parallel previously meant juggling many terminal tabs. Tracking "which tab does what" in your head was exhausting.

Agent View fixes that. Run claude agents and your terminal becomes a dashboard listing all running Claude Code sessions by status (running / waiting on user / completed).

Main features

• Session list: each row shows name, current activity, last update
• State grouping: "waiting on user" and "pinned" sessions float to the top
• Background send: type /bg in an existing session to stash it into Agent View
• New background launch: claude --bg "Process companies 50–100 in the prospect list" spawns without using the foreground
• Reattach: click a waiting session to drop back into normal chat

Example — sales scenario

Sales Claw has internally implemented parallel batches ("up to 3 awaiting approval"), and Agent View enables "monitor Sales Claw batches from outside":

# Split a 90-company target list into 3 parallel jobs
claude --bg "Sales Claw: process companies 1-30"
claude --bg "Sales Claw: process companies 31-60"
claude --bg "Sales Claw: process companies 61-90"

# Monitor all jobs from one screen
claude agents

90 companies in parallel with one screen for monitoring — operational burden is roughly 1/10 of "three terminals lined up visually managed." For implementation details, see Sales Claw Workflow overview. A follow-up article covers Claude Code 2.1.140's subagent_type / /goal fixes.

With Claude Code 2.1.139, run B2B outbound as policy-controlled autonomous operation.

無料・MIT ライセンス。インストールせずにライブデモも試せます。

無料でダウンロードライブデモを試すGitHub

3. Implementation for AI sales

Environment setup

# Upgrade Claude Code to 2.1.139 or newer
npm install -g @anthropic-ai/claude-code@latest

# Check version
claude --version

# Clone & start Sales Claw
git clone https://github.com/joseikininsight-hue/sales-claw-ts
cd sales-claw && npm install && npm start

Outbound flow with /goal

In the Sales Claw dashboard, click [Start Claude], then type into the PTY:

/goal Create 10 awaiting_approval entries. Process from the top of the target list.
Hand CAPTCHA to a human. Up to 60 minutes elapsed.

Monitor in real time via the "AI cost estimate" chip in the lower left of the dashboard.

Parallel operations with Agent View

# Split 300 companies into 3 jobs
claude --bg "Sales Claw process #1-100. /goal create 30 awaiting_approval"
claude --bg "Sales Claw process #101-200. /goal create 30 awaiting_approval"
claude --bg "Sales Claw process #201-300. /goal create 30 awaiting_approval"

# Monitor from one screen
claude agents

4. Cost estimate — the reality of 10,000 sends per month

Assumptions

• Targets: 10,000 companies/month (500/day × 20 business days)
• Judging model: Claude Haiku 4.5 (default for /goal)
• Copy worker: Claude Sonnet 4.6 (high-quality mode)
• Form fill: MCP Playwright (via Sonnet 4.6)
• FX: 1 USD = 150 JPY
• Exclusion: ~8% CAPTCHA / ~4% do-not-contact / ~6% no-form (Sales Claw internal values)
• Per-company tokens: ~4,500 in / ~1,200 out (internal values)

Cost breakdown

Comparison with outsourcing services

5. Risks of /goal in AI sales and safe-by-design

/goal is powerful, but for sales automation you must build legal / compliance / runaway risks into the design. Sales Claw defends against these by default.

Legal / compliance

• Japan's Specified Electronic Mail Act: 4 sender-info fields (company / name / contact / opt-out) must be in the body. Sales Claw auto-appends via preferences.complianceFooter: true
• Per-site terms: pages with "no sales" skip automatically
• No CAPTCHA bypass: stop into awaiting_approval on detect, hand off to a human
• Send log: every action recorded to action-log.json; auditable
• Opt-out: explicitly state "reply not needed" / "opt out" in copy

Three rules to prevent runaway

Pre-production checklist

6. Bottom line — into a new era of AI sales

By May 2026, AI sales automation has moved through four clear phases:

1. Phase 1: template sends (~2022)
2. Phase 2: AI copy generation (2023+)
3. Phase 3: AI + browser automation (2024+)
4. Phase 4: policy-controlled autonomous operation (from 2026-05-12) — AI continues until conditions are met; pre-send automatic checks and audit logs sit in between

With Claude Code 2.1.139, the "policy-controlled autonomous operation" philosophy Sales Claw has championed catches up to industry standard. Not unbounded auto-sending, but AI that continues while automated checks run pre-send and audit logs capture everything — that's the shape AI sales should take as of May 2026.

Next action: upgrade Claude Code → install Sales Claw → validate on 100 sample companies → run the first /goal loop. Start at Quick start and AI CLI setup. Sales Claw itself is free at the download page.

Read this. Now run.

無料・MIT ライセンス。インストールせずにライブデモも試せます。

無料でダウンロードライブデモを試すGitHub