The landscape of developer tools is shifting faster than ever. In just the past two weeks, we’ve seen major releases from Zed, Cursor, and Claude Code that collectively signal a clear trend: the conversation between developer and AI agent is becoming the primary development artifact, not the commit. Here’s a rundown of what’s new and why it matters.
Zed’s DeltaDB: Version Control Built for the Agent Era
On June 11, Zed’s founder Nathan Sobo announced DeltaDB, a new version control system designed from the ground up for AI-assisted development. The core premise is straightforward: Git captures snapshots at each commit, but the most important work happens between commits — in the conversation with your agent. DeltaDB captures every fine-grained operation and ties it to the conversation that produced it.
Instead of line-number-based references that break when code moves, DeltaDB anchors every reference to a stable delta identity. From any line in a past conversation, you can jump to the code as it stands now or as it stood the moment the agent wrote it. Agents can also draw on this history — they can pick up the context behind code they’re modifying and even query the prior agents that worked on it.
The implications for team collaboration are significant. A teammate can join a work session while the agent is still running, read the conversation, and annotate in real time — no commit, push, or pull request required. Zed is positioning DeltaDB as complementary to Git (which stays useful for CI checks and external integration), not a replacement. A beta is expected in a few weeks.
Cursor 3.7: Design Mode Gets Multi-Select and Voice
Cursor’s June 5 release focused on Design Mode, their browser-based UI editing feature. Two additions stand out. Multi-select lets you click multiple elements on the page simultaneously — the agent sees the selected elements, their code, the surrounding layout, and the visual relationships between them. You can then ask it to make one element match another, remove repeated content, or adjust a group of components at once.
Voice input works through the Design Mode overlay. The microphone stays available while the agent is mid-run, so you can queue the next change by voice without waiting for the current one to finish. For rapid UI iteration, this removes the keyboard entirely from the feedback loop.
Cursor SDK: Custom Tools, Auto-Review, and Subagent Nesting
The Cursor SDK update on June 4 brought the most developer-facing changes across both the TypeScript and Python SDKs. The headline feature is custom tools: you can now pass function definitions through local.customTools on Agent.create() or per send(), and the SDK exposes them to the agent through a built-in MCP server called custom-user-tools. Before this, exposing a custom capability meant standing up your own stdio or remote HTTP MCP server — now a function definition is enough, and it’s automatically visible to every subagent in the run.
For headless CI runs, the new local.autoReview setting replaces the default bypass-all-permissions behavior with a classifier that decides which tool calls to approve automatically and which to hold for review. You steer the classifier with natural-language instructions in permissions.json using autoRun.allow_instructions and autoRun.block_instructions:
{
"autoRun": {
"allow_instructions": [
"Read-only inspections of build artifacts under ./dist are fine."
],
"block_instructions": [
"Always pause delete operations so I get a chance to review them."
]
}
}Subagent nesting is also now supported to any depth, enabling complex multi-step workflows where a parent agent delegates specialized tasks to child agents, each keeping its own prompt and context.
Claude Code 2.1.170–2.1.172: Fable 5 and Sub-Agent Chains
Claude Code has been on a rapid release cadence. Version 2.1.170 (June 9) introduced Fable 5, a Mythos-class model Anthropic describes as exceeding the capabilities of any model they’ve made generally available. The same release fixed a bug where sessions launched from the VS Code integrated terminal weren’t saving transcripts.
Version 2.1.172 (June 10) unlocked sub-agent chains: agents can now spawn their own sub-agents up to 5 levels deep. This enables hierarchical delegation patterns — a planning agent dispatches a coding agent, which in turn dispatches a testing agent, all within a single Claude Code session. This release also fixed a critical bug where sessions using 1M context without usage credits would get permanently stuck — they now auto-compact back under the standard context limit.
Practical Takeaways
What ties these releases together is a shared bet: the editing experience is becoming a conversation experience. Zed is building a new version control layer for it. Cursor is making its SDK more programmable so you can define custom tools and permission boundaries. Claude Code is deepening sub-agent orchestration and model governance.
For day-to-day development, the most immediately useful changes are the Cursor SDK’s custom tools (which eliminate the need for external MCP servers for simple functions) and Claude Code’s sub-agent nesting (which makes complex multi-step tasks possible within a single session). DeltaDB is the most ambitious long-term bet — if it delivers on its promise of stable, conversation-anchored references, it could fundamentally change how teams collaborate on code.
These tools are all converging on the same conclusion: the commit is no longer the atomic unit of work. The conversation is. The tools that treat it as a first-class artifact will be the ones that stick around.