This week is packed with practical news for developers: Cursor transforms from editor to platform, Claude Code ships its biggest March update, Anthropic opens 1M context GA at standard pricing, and a trillion-parameter model appears… for free.

Highlight: Cursor Becomes an Agentic Platform#

Cursor had its biggest week ever — not one but three announcements that transform it from an AI code editor into a full platform.

Cursor v2.6 — Interactive UIs in Agent Chats (3/3)#

Agent chat can now render interactive UIs — charts from Amplitude, diagrams from Figma, whiteboards from tldraw. No more text-only responses.

Cursor Automations (3/5)#

The biggest news: always-on agents running 24/7 in the cloud, triggered by:

• Slack — “when there’s a bug report in #production, auto-create a draft fix”
• Linear/GitHub — “when an issue is assigned, research the codebase and propose an approach”
• PagerDuty — “when there’s an alert, check logs and create a summary”
• Webhooks — any custom trigger

Agents run in cloud sandboxes with memory that improves across runs. This is the leap from “AI assistant” to “AI teammate.”

JetBrains Support + Plugin Marketplace (3/4 and 3/11)#

• Cursor available on IntelliJ, PyCharm, WebStorm via Agent Client Protocol (ACP)
• 30+ new plugins from Atlassian, Datadog, GitLab, Hugging Face, PlanetScale
• Team marketplace for private plugins

TIP

If you’re a Java/.NET developer committed to JetBrains, you no longer have to choose between your familiar IDE and AI coding. Cursor comes to you.

Claude Code v2.1.76 — Biggest March Update#

Released March 14, this is the most significant update of the month:

MCP Elicitation Support#

MCP servers can now request structured input mid-task via interactive dialogs. Example: a deployment MCP server can ask “Deploy to staging or production?” while a task is running. Comes with new hooks: Elicitation and ElicitationResult.

New Features#

Critical Bug Fixes#

• RTL text: Hebrew, Arabic render correctly on Windows Terminal, conhost, VS Code
• LSP servers: Working again on Windows (was broken due to malformed file URIs)
• Voice mode: Fixed activation on fresh installs
• Plugin marketplace: Fixed git submodules, @ref separator

NOTE

worktree.sparsePaths is a hidden gem for monorepo users. Instead of Claude Code scanning your entire 10GB repo, you declare only the folders you need.

1M Context GA + Web Search GA — Standard Pricing#

On March 13, Anthropic announced two major changes:

1. 1M token context window GA for both Opus 4.6 and Sonnet 4.6 — no long-context surcharge. Previously required a beta header or extra usage.
2. Web search tool and programmatic tool calling officially GA — no beta header needed.

This means: every developer using the Claude API now gets 1M context at standard pricing. 1 million tokens ≈ ~750,000 words ≈ the entire source code of many large projects.

Anthropic Launches Claude Partner Network — $100M#

On March 12, Anthropic announced the Claude Partner Network with a $100 million 2026 commitment:

• Partners: Accenture (30,000 professionals in training), Deloitte (470,000 employees get access), Cognizant, Infosys
• First certification: Claude Certified Architect, Foundations — technical exam for solution architects
• Partner-facing team scaled 5x
• Applied AI engineers join partners on live deals

TIP

When Accenture and Deloitte train tens of thousands on Claude, demand for developers who build on Claude API/Claude Code will surge. Claude Certified Architect is a credential worth considering if you’re going the AI engineering route.

Same Day: The Anthropic Institute#

Anthropic established The Anthropic Institute led by co-founder Jack Clark, merging three groups: Frontier Red Team, Societal Impacts, and Economic Research. Notable: a new report shows 75% of programming tasks can now be AI-assisted, but there’s no evidence of systematic unemployment increases — only declining entry-level hiring in high AI-exposure fields.

Anthropic Doubles Usage Limits — Off-Peak#

From March 13 to March 27, usage limits are doubled outside 8 AM - 2 PM ET for all plans.

TIP

In Asia-Pacific time zones, the off-peak window (2 PM - 8 AM ET) covers most of your working hours. Take advantage while it lasts.

OpenAI Codex Desktop Comes to Windows#

After 500,000+ developers on the waitlist, the Codex desktop app officially launches on Windows (Microsoft Store). Features:

• Manage multiple agents in parallel
• Built-in worktree support and Git integration
• Custom native agent sandbox for security
• Temporarily free for ChatGPT Free and Go plans

GitHub Copilot — JetBrains Agentic GA#

On March 11, GitHub shipped agentic features GA for JetBrains:

• Custom agents, sub-agents, plan agent
• Auto-approve for MCP support
• CLI v1.0.5 with new /extensions command

WARNING

Copilot Student plan restructured (3/12): GPT-5.4 and Claude Opus no longer individually selectable (still available in Auto mode). Students can upgrade to Pro/Pro+ while keeping Student Pack benefits.

Hunter Alpha — 1 Trillion Parameter Model, Free#

On March 11, Hunter Alpha appeared on OpenRouter — a stealth model with:

• 1 trillion parameters — largest ever offered publicly
• 1M token context window
• Designed for agentic use: long-horizon planning, multi-step execution
• $0/M tokens (prompts are logged to improve the model)
• Speed: ~48 tokens/s

First time developers can try a 1T-param model for free — though the team behind it is unclear and real-world benchmarks need more evaluation time.

Tools & Releases#

Gemini 3.1 Flash-Lite#

Google’s efficiency-focused model: 2.5x faster, 45% faster output, $0.25/1M input tokens — cheapest in the Gemini family. Great for classification, extraction, simple Q&A.

SurePath AI — MCP Policy Controls#

MCP reached 97 million monthly SDK downloads. SurePath AI launched controls for organizations to manage in real-time which MCP servers are permitted — critical for enterprise adoption.

Benchmark Watch#

NOTE

GPT-5.4 scored only 57.7% on SWE-bench — far behind Claude and Gemini for coding. However, GPT-5.4 excels at general professional work (83% GDPVal). Pick your model by task: Opus/Gemini for coding, GPT-5.4 for general knowledge work.

Morgan Stanley: AI Breakthrough Incoming#

On March 13, Morgan Stanley warned that a transformative AI leap will happen in H1 2026. Compute scaling still works — 10x compute ≈ 2x intelligence. CEOs are already executing large-scale workforce reductions.

WARNING

Morgan Stanley is one of Wall Street’s most conservative institutions. Developers should pay attention — not to be afraid, but to prepare the right skills.

Meta: Delays Avocado, Acquires Moltbook#

• Model Avocado is delayed to May — falls short of Gemini 3.0, Claude, GPT-5.4. Meta is reportedly discussing licensing Gemini from Google temporarily.
• Meta acquired Moltbook — a social network for AI agents, bringing the team into Meta Superintelligence Labs (MSL). Also announced plans for 4 generations of custom AI chips (MTIA 300-500) before end of 2027.

Community Pulse#

• 85% of developers regularly use AI tools, 42% of new code is AI-assisted
• 57% of organizations have AI agents in production
• GDC 2026 wrapped March 13 — AI dominated game dev conversations
• Claude Code grew from 4% (May 2025) to 63% (Feb 2026) — the most popular AI coding tool

Quick Takes#

• VS Code 1.111: Agent permissions, Autopilot preview — VS Code becoming an agentic IDE
• Slack MCP Server: AI agents connect directly to Slack workspaces
• NIST AI Agent Security Framework: Comment period closed March 9. Final framework expected in Q2 2026.
• Google ships Groundsource — transforms news into structured data

What to Watch Next Week#

• NVIDIA GTC 2026 (March 16-19) — Jensen Huang keynote, NemoClaw (open-source enterprise AI agents), new chips
• Meta Avocado — more on the Gemini licensing deal?
• Claude Code — agent improvements likely shipping
• MCP 2026 Roadmap — Working Groups actively discussing

What new feature did you try this week? Cursor Automations, Claude Code 1M context, or Codex Desktop? Share below.