This is a living signal. The body reflects the current state of Claude Code as a platform. The What’s New section is updated when something material ships.
What’s New This Week
Claude Code Channels launched in research preview on March 20, 2026. The feature lets external systems push events directly into a running Claude Code session via MCP – CI results, monitoring alerts, Telegram messages. Claude reads the event and reacts, even when you’ve stepped away from the terminal. Requires v2.1.80+ and a claude.ai account. Telegram and Discord are the two official integrations; a reference API supports custom MCP channels. Two-way reply is included: Claude responds back through the same channel.
For a full breakdown of the architecture and what it enables, see Claude Code Channels: The Away Problem, Solved.
Changelog
| Date | Summary |
|---|---|
| 20 Mar 2026 | Signal launched. Channels research preview ships; competitive context with Codex Astral acquisition. |
The Platform Today
Claude Code launched in early 2025 as a command-line coding assistant. The pitch was simple: point it at your codebase, ask it to build something, watch it work. It reads files, writes code, runs shell commands. The starting point was a sophisticated autocomplete that could navigate a whole repo rather than a single file.
That framing is already out of date.
By March 2026, Claude Code has accumulated enough distinct capabilities that “coding assistant” undersells it. The current stack:
Terminal CLI and IDE integrations. Claude Code runs in your terminal natively on macOS, Linux, and Windows. VS Code integration is available. The same model powers a browser-based interface and desktop app for users who prefer it.
AGENTS.md architecture. The persistent configuration layer. A markdown file in your repo root – or per-directory – that tells Claude what it’s working with, what tools are available, what conventions apply, and how it should behave. It’s not a prompt. It’s infrastructure. Agents carry context across sessions because the configuration is committed to the repo. There’s a full breakdown of what AGENTS.md enables here.
MCP ecosystem. Model Context Protocol is the standardised integration layer. MCP servers extend what Claude Code can do – file system, shell, web search, database queries, API calls, external services. The extension surface is everything that can implement the protocol, which is increasingly everything. This is the action surface that makes Claude Code useful beyond code editing.
1M context window. Shipped with Claude Opus 4.6 and Sonnet 4.6. A million tokens is roughly 750,000 words – enough to hold an entire medium-sized codebase in context simultaneously. The practical implication: Claude Code can reason about the whole system, not just the files you’ve pointed it at. Cross-repository refactors, whole-codebase audits, understanding the full dependency graph before making changes. The limitations that required breaking work into smaller sessions are largely gone.
Channels. Launched today in research preview. Push-based event model: external systems send events into a running session, Claude reacts, replies back via the same channel. This is the event loop that was missing. More on this below.
Why This Signal Exists
Claude Code is the most architecturally interesting of the current generation of coding agents. Not necessarily the most used – Codex has 2M weekly active users, Copilot has the deepest enterprise penetration, Cursor has strong developer loyalty. But the architectural choices Anthropic is making with Claude Code have implications beyond the tool itself.
The combination of AGENTS.md (persistent config), MCP (extensible action surface), 1M context (reasoning scope), and Channels (event model) is not just feature additions. Each of these is a primitive. Put them together and you get something that looks less like a coding assistant and more like an autonomous agent runtime that happens to be good at writing code.
That’s worth tracking. The questions this signal is designed to answer:
- Where does the AGENTS.md architecture go? Is it a file format or is it becoming a protocol?
- How does the MCP ecosystem consolidate? Which servers become infrastructure?
- Does the Channels model generalise – and what does a mature version of that look like?
- Can Claude Code hold its position against Codex’s distribution advantage and Copilot’s enterprise reach?
- What do the security and governance implications look like as these agents gain more autonomy?
The broader thesis – that we’re in the middle of an agentic turn where AI moves from tool to operator – is developed further at russellclare.com/ai-agentic-turn/. Claude Code is one of the clearest examples of that transition happening in real time, which is why it warrants ongoing tracking.
The Architecture
Four primitives. Here’s how they fit together.
AGENTS.md: configuration as infrastructure
The conventional way to configure an AI assistant is through the prompt. You write instructions, the model follows them, the instructions disappear when the session ends. AGENTS.md is different: configuration lives in the repo, versioned alongside the code, shared across the team, persistent across sessions.
The implications compound. When configuration is in the repo, every engineer using Claude Code on the same project gets consistent behaviour. When it’s versioned, you can track how agent behaviour has evolved. When it’s per-directory, you can have different behaviour in different parts of the codebase – different conventions for frontend vs backend vs infrastructure.
It’s a small shift from “instructions I type before I start” to “configuration I commit once.” But it changes the unit of work. You’re not configuring a session; you’re configuring an agent that works on your codebase.
MCP: the extensible action surface
Model Context Protocol standardises how tools connect to agents. Rather than building custom integrations for every tool, MCP provides a shared protocol. Claude Code ships with core capabilities – file system, shell, web search – and MCP extends that surface to anything with an MCP server implementation.
The ecosystem is maturing quickly. Database clients, API wrappers, monitoring system integrations, version control operations beyond what git provides natively. The action surface is expanding faster than any single team could ship first-party.
The competitive parallel is obvious: MCP is doing for agent tooling what npm did for JavaScript libraries. The question is whether it stays Anthropic-centric or becomes genuinely cross-agent infrastructure.
1M context: reasoning at the system level
The practical change from a 200K context window to 1M is not just “more code fits.” It’s that the unit of reasoning changes. You stop thinking about which files to include and start thinking about what you’re trying to accomplish. The model can hold the whole system.
For refactors that span many files, for understanding how a change in one part affects behaviour elsewhere, for auditing security properties across an entire codebase – the 1M window makes these workflows tractable in a single session in a way they weren’t before.
Channels: the event loop
This is the newest piece. A running Claude Code session can now receive external events and react to them – no human relay required. CI failure at 2am, monitoring alert on Saturday, new PR opened while you’re in a meeting. Claude gets the event, acts on it, reports back via Telegram or Discord.
The architectural significance: this transforms Claude Code from a synchronous tool (you type, it responds) to an asynchronous agent (the world sends events, it reacts). That’s the last piece needed for a genuinely autonomous deployment pattern. Session running in tmux, channels connected, event sources configured. The agent handles incidents. You check your messages.
Channels is in research preview now. The production version of this – with better security primitives, broader channel support, enterprise-grade audit logging – is the thing to watch.
The Competitive Context
The developer AI tooling space is consolidating fast. Four positions that matter:
Claude Code (Anthropic): Agent in the terminal, AGENTS.md architecture, MCP ecosystem, 1M context, Channels. Strong on architecture, growing on adoption. No published WAU numbers.
OpenAI Codex: 2M weekly active users as of March 2026. 3x user growth and 5x usage increase since January. On March 19, OpenAI acquired Astral – the company behind uv, Ruff, and ty, the modern Python toolchain with 126M+ monthly downloads for uv alone. Codex now owns Python dependency management and linting at scale. Distribution advantage is significant; the Astral acquisition adds toolchain depth that goes beyond the model itself. The implications of the Astral acquisition are covered here.
OpenClaw: Open-source agent framework, cited by Jensen Huang as the fastest-growing OSS project. The platform layer rather than the model layer – OpenClaw provides the harness that agents run inside. NemoClaw, NVIDIA’s enterprise security stack, is built on top of it. This is the infrastructure position: if Claude Code or Codex is the application, OpenClaw is competing to be the runtime. NemoClaw and the enterprise security layer.
Cursor and GitHub Copilot: Cursor has strong developer loyalty through tight IDE integration and good UX. Copilot has the deepest enterprise penetration by distribution – it comes with GitHub, and GitHub is where most enterprise code lives. Neither is building toward autonomous agent workflows in the way Claude Code and Codex are; both remain closer to the “AI-assisted coding” framing.
The interesting competitive tension is between Claude Code and Codex specifically. Both are moving toward autonomous agent architectures. Both are building platform layers – Codex through toolchain ownership (Astral), Claude Code through protocol standardisation (MCP) and configuration infrastructure (AGENTS.md). They have different distribution bases and different architectural bets. One of them probably ends up as the dominant platform for agentic development workflows. We don’t know which yet.
The framework race – OpenClaw’s position – is a separate bet. If the model layer commoditises and the value shifts to the harness, OpenClaw’s open-source position becomes very interesting. If models stay differentiated, the framework matters less.
What to Watch
Specific developments that would update this signal’s thesis:
AGENTS.md standardisation. If AGENTS.md becomes a cross-tool standard – adopted by Codex, Cursor, or Copilot – that changes the competitive picture significantly. If it stays Anthropic-specific, it’s a retention mechanism rather than an ecosystem. Watch for third-party adoption.
MCP ecosystem consolidation. Which MCP servers become the defaults? If a small number of high-quality servers become infrastructure that every agent relies on, that’s a platform layer forming. Watch for adoption signals in the MCP server ecosystem.
Channels in production. The research preview has real constraints: claude.ai login, enterprise opt-in, limited integrations. When Channels moves to GA and the security model matures, that’s the signal that Anthropic is serious about the always-on deployment pattern. Watch for the enterprise audit logging story specifically.
Codex toolchain integration depth. The Astral acquisition gives Codex uv and Ruff. How deeply integrated does that get? If Codex becomes meaningfully better at Python-heavy workflows because it owns the toolchain, that’s a durable advantage that’s hard to match.
Claude Code adoption numbers. Anthropic hasn’t published WAU figures. When they do – or when credible third-party estimates emerge – that will clarify whether the architectural lead translates to adoption.
Enterprise governance stories. NemoClaw is building the security layer on OpenClaw. Who builds the equivalent for Claude Code? The gap between “interesting for individual engineers” and “deployable in an enterprise” is largely a governance and audit story.
Security incidents. The attack surface here is real. Channels, MCP servers with broad permissions, 1M context windows that can be poisoned with adversarial content – prompt injection at agent scale is an unsolved problem. The first high-profile incident will accelerate governance tooling development fast.