🌟 Vasilij’s Note
This week crystallised something I keep coming back to with clients. The Claude Code updates landing in March - voice mode, /loop, Channels, Computer Use - aren't interesting because of the feature list. They're interesting because, taken together, they remove the last plausible excuse for staying in reactive mode. You can now instruct Claude Code from your phone via WhatsApp while you're in a client meeting, have it run a background reviewing action, do research, and return to a completed output. The infrastructure is no longer the constraint. Deciding which workflow to hand over first is and developing the trust in AI tools.
In Today's Edition:
This Week in Agents | What Changed
Claude Code ships Channels - Discord and Telegram control of your coding agent. On March 20, Anthropic released Claude Code Channels as a research preview. Developers can now connect Claude Code to Discord or Telegram via the MCP plugin, message it asynchronously from anywhere, and receive notifications when tasks are completed. This moves the interaction model from synchronous "ask-and-wait" to asynchronous autonomous partnership. Requires Claude Code v2.1.80+ and the Bun runtime. Official setup documentation is live.
Voice mode, /loop, and Opus 4.6 default - the full March release in one line. Claude Code v2.1.76 shipped push-to-talk voice mode (20 languages, spacebar to speak), the /loop command for scheduled recurring tasks, Opus 4.6 as the default model with 64K output tokens and 1M context on Max/Team/Enterprise plans, and context compaction for ultra-long sessions. The /effort command simplifies reasoning depth to three levels (low/medium/high), with "ultrathink" as a temporary override.
Google Gemini agents now crawling the dark web at 10 million posts per day. Google Threat Intelligence launched a public preview of Gemini-powered dark web monitoring on March 23. The service processes up to 10 million daily posts from criminal forums, initial access broker listings, and data leak sites, distils them into organisation-specific alerts, and claims 98% accuracy in internal testing. Setup takes minutes: the system builds a profile from open-source data, then auto-generates a week of back-filled alerts. Customers receive prioritised threat classifications, not raw data.
Top Moves - Signal → Impact
Bezos is raising $100B for an AI-driven manufacturing acquisition fund. Jeff Bezos is in early discussions with major asset managers - including meetings in the Middle East and Singapore - to raise $100 billion for a fund targeting manufacturing companies in chipmaking, defence, and aerospace for AI-accelerated automation. This is not an AI startup fund. It is a bet that AI-driven operational transformation of the physical industry is the next decade's defining arbitrage.
→ The capital allocation signal here is significant for consultancies advising industrial clients. AI operational transformation in regulated, capital-intensive sectors is moving from pilot to investment thesis.
OpenAI acquires Astral - Python tooling (uv, ruff, ty) moves inside ChatGPT's parent. OpenAI has acquired Astral, maintainer of the Python toolchain tools uv and ruff that have become standard in agentic coding stacks. Community response is mixed: the tools can be forked, but long-term governance uncertainty is real for any team that has standardised on them.
→ If your Claude Code or agent build stack uses uv or ruff - and most do - monitor for roadmap divergence. Open-source tools acquired by AI labs have a track record of shifting toward proprietary integration.
Nudge Security releases AI agent discovery tools - 80% of organisations already at risk. On March 24, Nudge Security launched tooling that allows IT and security teams to discover every AI agent their employees have created across platforms, including Microsoft Copilot Studio, identify what data each agent can reach, and flag configurations that violate security policy. The warning accompanying the release: 80% of organisations are already seeing risks from AI agents with excessive data access, and most were created without IT involvement.
→ Shadow AI agent sprawl is no longer a theoretical risk. If your firm or your clients are on Microsoft 365, agents are being created right now that IT does not know about and cannot audit. The governance conversation is not a future agenda item.
Upskilling Spotlight | Learn This Week
Claude Code Academy - free, official, practical.
Anthropic's Academy courses covering Claude Code are available at no cost, including the architecture behind /loop, skills, MCP elicitation, and the Channels setup. The MCP elicitation module is the most immediately applicable: it covers how to build MCP servers that request structured input mid-task via interactive forms without interrupting the workflow - the pattern that powers most reliable agent deployments.
Maker Note | What I built this week
This week, I recorded a full breakdown of the ten Claude Code features most teams are leaving on the table - billing traps, context management, hooks, MCP integrations, and Agent Teams.
Decision: the gap isn't capability, it's configuration. One developer's 201 sessions would have cost $5,623 at API rates. On Max, that's $100 a month. The billing audit alone is worth 30 seconds of your time before you read anything else this week.
Anthropic ships between 60 and 100 updates to Claude Code every single day. If you set it up a few months ago and haven't touched your config since, you're running a completely different tool to what's available now.
Operator’s Picks | Tools To Try
Claude Code /loop with cron scheduling
Use for recurring automated checks: PR review, deployment monitoring, data validation cycles, compliance checks on a defined schedule. Pairs cleanly with n8n or webhook output for team notification.
Caveat: Behaviour on ambiguous criteria degrades quickly. Write explicit acceptance criteria before deploying to any client-facing workflow.
Cursor Composer 2
Use for frontend development and complex multi-file coding workflows where Claude Code's terminal-first approach is not the right fit. Standard model at $0.50/M input - materially cheaper than Opus 4.6 for tasks where reasoning depth isn't the constraint.
Pair with: Claude Code for codebase-level context; Composer 2 for rapid iteration on specific components.
Deep Dive | Thesis & Playbook
Claude Code in March 2026: What the Feature Stack Actually Changes for Delivery Teams
The framing that Jensen Huang used at GTC - ChatGPT, o1, then Claude Code as the three chapters of the AI era - wasn't a compliment. It was a capability classification. Claude Code is not a better autocomplete. It is the first tool in this category that can sustain a multi-hour, multi-file, multi-application workflow without a developer managing each step. The March releases operationalise that claim.
On Paper
Voice mode ships push-to-talk with 20 languages and a controlled, non-invasive interaction model - hold spacebar to speak, release to send. Not ambient surveillance, not always-on. A deliberate input method that removes the terminal as a physical requirement.
/loop creates cron-style scheduled tasks inside a Claude Code session. PR review on a 30-minute cycle, deployment health checks, data validation on a defined schedule. These run without a human at the keyboard.
Channels connect Claude Code to Discord and Telegram via MCP. Send a task from your phone, receive results when complete. The session runs on your machine; only chat messages transit the network.
1M context on Max, Team, and Enterprise plans means entire codebases fit inside a single session. Claude Opus 4.6 at 64K default output and 128K ceiling handles the responses.
In Practice
/loop performs reliably on rule-based, binary checks. It degrades on subjective criteria the same way any automation degrades on subjective criteria. The discipline required is identical to any agent deployment: write the acceptance standard explicitly before deploying, not after encountering bad output.
Channels require setup time: v2.1.80+, Bun runtime, bot configuration, plugin install, and token pairing. Budget 20-30 minutes for a working Telegram integration. It is not plug-and-play for non-technical team members. This is a developer-facing feature that enables non-developer oversight, not non-developer operation.
Voice mode transcription accuracy is notably improved for technical terms - the March notes specifically cite regex, OAuth, and JSON as improved. The improvement matters for dictating complex requirements in developer contexts.
The 1M context window is currently in beta. For teams running large monorepo workflows, the new worktree.sparsePaths setting (March 10) allows checking out only the directories needed via git sparse-checkout - critical for managing context on large codebases without burning tokens on irrelevant files.
Issues / Backlash
Context compaction is automatic in long sessions, which is useful for coherence but means earlier context is summarised, not preserved verbatim. For compliance-sensitive workflows or audit requirements, document session boundaries explicitly.
The Channels architecture requires the Bun runtime, which is not universally present in enterprise environments. An IT security review of the MCP plugin model will be required before production deployment in any regulated firm.
Computer Use remains a research priority. It is available for testing but should not be in client-facing production workflows without defined fallback and approval procedures.
My Take (What to do)
Startup (15–40 staff):
The /loop and Channels combination is your highest-value unlock this month. Identify one recurring check your team does manually more than three times weekly - PR quality review, deployment verification, client report generation. Configure /loop to run against explicitly written criteria. Route results to Slack via n8n webhook. Measure actual time saved over four weeks before adding a second loop. Channels are secondary: set it up for one developer to validate the setup pattern before distributing to the team.
SMB (50–120 staff):
You have a process documentation problem that makes /loop noisy. Before configuring any scheduled task, your delivery lead needs to document the acceptance criteria for that task in a format that a new team member could follow without asking questions. If that documentation doesn't exist, write it first - it is valuable independent of automation. Once it exists, the loop configuration takes 30 minutes. Focus on three high-frequency workflows: PR review, status report generation, and client data validation. Assign one ops team member as the Claude Code owner who manages the plugin registry, reviews loop outputs weekly, and owns the governance record.
Enterprise (150–250 staff):
The competitive pressure from GPT-5.4 and Cursor Composer 2 is real. Your clients' development teams are evaluating both. You need a clear internal position on where each tool fits and why - not a preference, a decision framework grounded in actual capability testing. Run a structured evaluation of Claude Code Channels vs GPT-5.4 computer use against two or three of your actual client delivery workflows this month. Document the results. The decision to standardise matters less than having a documented rationale. On governance: the Channels architecture (MCP plugin, device-local execution, network-transited chat only) is materially more auditable than alternatives that route through third-party infrastructure. That is a compliance advantage worth surfacing in client conversations.
How to Try (15-minute path)
Update Claude Code:
npm update -g @anthropic-ai/claude-code. Verify version 2.1.76 or later withclaude --version. (2 min)Run
/loopon a simple test: ask Claude to check every 5 minutes whether a specific file in a test repo meets three explicit criteria you write out in plain English. Observe what it does when criteria are met and when they aren't. (8 min)Document what broke or produced unexpected output. That gap between your expectation and the loop's behaviour is exactly what needs to be specified before this runs on a client engagement. (5 min)
Success metric: a written acceptance criteria document for one real workflow that is specific enough for /loop to evaluate unambiguously. If you can't write it in 30 minutes, the workflow isn't ready for automation - and that's useful information.
Spotlight Tool | Claude Code Channels
Purpose: asynchronous control of Claude Code sessions via Telegram or Discord - for developers who cannot or should not be at a terminal during active agent sessions.
Edge: MCP-native architecture. No proprietary relay. The session executes locally on your machine; chat messages only transit the network. More auditable than alternatives routing through third-party infrastructure.
→ Asynchronous task delegation from mobile or remote location
→ Push notifications when long-running tasks complete
→ MCP plugin model - deploy once across team via plugin marketplace
→ Compatible with /loop for delegated recurring task monitoring
→ Research preview: stable enough to test, not yet production-hardened
Try it: code.claude.com
What did you think of today's email?
Sponsored - Partner
n8n – An open‑source automation platform that lets you chain tools like DeepSeek, OpenAI, Gemini and your existing SaaS into real business workflows without paying per step. Ideal as the backbone for your first serious AI automations. Try: n8n
Did you find it useful? Or have questions? Please drop me a note. I respond to all emails. Simply reply to the newsletter or write to [email protected]