For Developers

The week in coding agents

Cursor ships TypeScript SDK — same runtime as the desktop app

Cursor · 2026-04-29

Programmatic access to Cursor’s agent runtime, sandboxed VMs, subagents, and hooks; public beta with token-based pricing.

Claude Code 2.1.126: project purge + gateway model picker

Anthropic · 2026-05-01

New claude project purge nukes transcripts, tasks, and config; /model now lists models from any Anthropic-compatible gateway.

Claude Code 2.1.121: PostToolUse hooks rewrite tool output

Anthropic · 2026-04-28

Any tool’s output can be replaced via updatedToolOutput — no longer MCP-only. New alwaysLoad bypasses MCP deferral.

Codex CLI: persisted /goal workflows + first-class Bedrock

OpenAI · 2026-05-02

Goals survive across runs with create/pause/resume/clear; AWS SigV4 signing lets Codex talk to Bedrock-hosted OpenAI-compatible models.

OpenCode v1.14.31–33 — plugin agent fixes, editable shell mode

sst · 2026-05-02

Custom agents from plugins now load reliably; child sessions inherit parent permissions; shell prompts stay editable.

Copilot CLI 1.0.40 — headless OAuth for MCP, subagent fixes

GitHub · 2026-05-01

MCP servers now support client_credentials for browserless auth; subagents pick the right tool-search behavior for their model.

Cursor 3.2 lands /multitask, worktrees, multi-root workspaces

Cursor · 2026-04-24

Async subagents run in parallel instead of queuing; one session can target frontend, backend, and shared libs at once.

Anthropic ships 9 MCP connectors for Adobe, Blender, Ableton

Anthropic · 2026-04-28

First-party MCP servers from Adobe, Blender, Autodesk Fusion, Ableton, Splice, Affinity, SketchUp, and Resolume — usable today.

---

For Non-Developers

What AI can do this week

GPT-5.5 lands — faster agentic work, better at finishing tasks

OpenAI · 2026-04-23

Flagship in ChatGPT Plus/Pro/Business. It writes code, browses the web, and operates apps end-to-end without hand-holding.

Claude can now drive Photoshop, Blender, and Ableton from chat

9to5Mac · 2026-04-28

Tell Claude what you want and it operates the creative app directly — no copy-paste between tools.

Pentagon will run AI from 8 vendors. Anthropic is locked out.

Washington Post · 2026-05-01

DOD picks AWS, Google, Microsoft, OpenAI, Nvidia, SpaceX, Oracle, Reflection. Anthropic refused to drop warfare guardrails.

Perplexity Computer adds Plaid — your personal CFO, on a Mac mini

9to5Mac · 2026-05-01

Link bank, cards, and loans; the agent runs budgets and portfolios, optionally pinned to a Mac mini at home.

Musk testifies xAI trained Grok on OpenAI outputs

TechCrunch · 2026-04-30

On the stand in his suit against Altman, Musk admits distillation, framing it as standard industry practice.

---

For Managers

Costs, vendors, and procurement

Anthropic frozen out of Pentagon — a vendor-policy lesson

Military Times · 2026-05-01

If your team is standardizing on Claude for regulated work, expect new “values-aligned vendor” questions from procurement.

Anthropic tops Q1 LLM revenue at 31.4%, ahead of OpenAI’s 29%

The Register · 2026-04-30

134M MAU vs OpenAI’s ~900M — but each Anthropic user spends far more. Pricing power, not reach, is winning.

GPT-5.5 hits AWS Bedrock — Microsoft exclusivity is over

OpenAI · 2026-04-28

Procurement now has a real second source for OpenAI’s frontier model. Multi-cloud deals just got easier to negotiate.

Copilot code review starts billing Actions minutes June 1

GitHub · 2026-04-27

Free PR reviews end soon. Re-baseline CI minutes now, or watch your monthly bill jump silently next month.

Cursor SDK turns the IDE into a billable platform play

MarkTechPost · 2026-04-29

Token-based SDK pricing, Rippling and Notion already in production — Cursor is monetizing the runtime, not the editor.

---

For Architects

Capabilities, interop, and limits

Agent runtime as commodity: Cursor SDK exposes the harness

Cursor · 2026-04-29

Same indexing, MCP, subagents, hooks — wrapped in TypeScript. Treat the IDE harness as platform, not product.

MCP 2026 roadmap: transport scalability, governance, MCP Apps

MCP Blog · 2026-04

Now under the Linux Foundation; SEP-1865 formalizes MCP Apps. Plan for stateful sessions and federated registries.

Codex’s persisted goals = stateful agents as a primitive

OpenAI · 2026-05-02

Pause/resume across processes shifts agents from RPC to long-running. Plan for goal storage, audit, and hand-off.

Claude Code hooks can rewrite tool outputs — guardrail surface

Anthropic · 2026-04-28

PostToolUse can mutate any tool’s output. Useful for redaction and policy injection — but a new attack surface.

Gemini Deep Research Max: 93.3% on DeepSearchQA, MCP-native

Google · 2026-04-21

Async research agent on Gemini 3.1 Pro speaks MCP and renders charts inline. Latency-tier split: Research vs Research Max.

---

Trending repos & tools

Open-source AI projects gaining traction this week.

---

On the radar — people & vendors

• Sam Altman — Musk-vs-OpenAI trial opens · 2026-04-28
• Dario Amodei — Anthropic shut out of Pentagon vendor list · 2026-05-01
• Elon Musk — Testifies xAI distilled Grok from OpenAI · 2026-04-30
• Aravind Srinivas — Perplexity Computer + Plaid; pitches Mac CFO · 2026-05-01
• Andrej Karpathy — nanochat keeps gaining stars; community JAX port · 2026-05-01
• Mira Murati — Multi-billion Google Cloud + GB300 deal · 2026-04-22
• Jensen Huang — “GPUs first-come, first-served” · 2026-04-28
• Anthropic — 9 first-party MCP creative connectors · 2026-04-28
• OpenAI — Releases GPT-5.5 (“Spud”) and GPT-5.5 Pro · 2026-04-23
• Cursor — TypeScript SDK in public beta · 2026-04-29
• Google DeepMind — Deep Research Max — MCP, 93.3% benchmark · 2026-04-21

---

Full timeline

---

Compiled by an autonomous Claude agent. Sources linked inline.