This article summarizes notable feature releases across major AI coding agents in mid-2026 — primarily Claude Code (Week 25/26, early July) and Cursor (multiple June releases). Aider updated model support without architectural changes. A gap analysis identifies features that remain unaddressed by existing tools and those at parity with recent implementations.
Key Points
Claude Code introduced shell mode that explains command failure output, and /rewind that restores a cleared conversation (going beyond earlier passthrough/snapshot-only implementations).
Claude Code added background subagent permission prompts surfaced in the main session, /cd for changing working directory mid-session, sub‑agent nesting (5‑level cap), Tool(param:value) permission rules, /config settable from -p mode, and Artifacts for live shareable session pages (Team/Enterprise beta).
Cursor released v3.8 Automations (event‑triggered workflows via /automate, e.g. PR review on Slack reaction) and v3.9 Plugin System (portable .cursor-plugin/ directory bundling skills, rules, MCP servers, commands, hooks; managed at user/team/workspace level).
Cursor’s SDK exposed custom tools through a built-in MCP server (custom-user-tools), a natural‑language‑steered permission classifier (allow/block rules in permissions.json), and unlimited subagent nesting.
Aider focused on model support (GPT‑5 family, Grok‑4, Kimi K2).
Concepts
Shell Mode – A Claude Code mode (! npm test) that feeds command output back into the conversation and prompts the model to explain failures or take action without a second user prompt.
/rewind – Restores a pre-/clear conversation in Claude Code, going beyond a single snapshot restore by resuming the conversation history.
Sub‑agent nesting – Ability for one agent to spawn another. Claude Code caps the chain at 5 levels; Cursor allows arbitrary depth.
Background Agents – Sub‑agents that previously auto‑denied permission prompts; now Surface those prompts in the main Claude Code session.
/cd – Changes the working directory mid‑session without rebuilding the prompt cache.
Tool(param:value) – Claude Code permission rules specifying allowed tools and parameter constraints (e.g., Agent(model:opus)).
/config – Sets configuration variables from command‑line -p mode.
Artifacts – Live, shareable session pages in Claude Code (Team/Enterprise beta).
Automations – Cursor’s event‑triggered workflows defined with /automate (e.g., review PR on Slack reaction, update dependencies). Triggered by webhooks, Slack reactions, etc.
Plugin System – Cursor’s portable plugin bundle: a single .cursor-plugin/ directory with plugin.json, skills, rules, MCP servers, commands, and hooks. Managed at user, team, or workspace level.
Custom Tools (SDK) – Cursor SDK allows defining custom function tools exposed via a built-in MCP server (custom-user-tools), visible to all sub‑agents.
Auto‑Review Classifier – Cursor’s natural‑language steered allow/block rules in permissions.json to control what the agent automatically reviews.
Details
Claude Code Week 25/26 (July 2026)
The feature set converged with earlier experimental implementations but added key improvements:
Shell Mode
Invoked with ! <command>.
Differs from earlier passthrough (which was zero‑token passthrough only): the model now receives command output and proactively explains errors, suggests fixes, or takes corrective actions.
/rewind
Restores the conversation to the state before the last /clear.
Differs from earlier snapshot‑only restore: it resumes the conversation, not just a one‑off snapshot.
Background Subagent Prompts
Previously, background sub‑agents would auto‑deny permission prompts. They now surface those prompts in the main session, allowing the user to approve.
/cd
Changes the session’s working directory mid‑conversation without invalidating or rebuilding prompt cache.
Sub‑agent Nesting
Agents can spawn sub‑agents. Claude Code enforces a maximum chain depth of 5 levels.
Tool Permission Rules
New syntax Tool(param:value) allows precise permission rules for agent tools, e.g., Agent(model:opus) restricts the agent to a specific model.
/config
Configuration variables can be set from -p mode (non‑interactive) using /config key=value.
Three major releases in one month introduced two broad new capabilities: Automations and a Plugin System.
v3.8 Automations
/automate command to define event‑triggered workflows.
Triggers include Slack reactions, webhooks, and other events.
Example: automatically review a PR when a Slack reaction is added, or update dependencies on a schedule.
v3.9 Plugin System
Portable, versioned directory: .cursor-plugin/ containing plugin.json plus subdirectories for skills, rules, MCP servers, commands, and hooks.
Managed at three levels: user, team, or workspace.
Bundles everything needed for a custom extension in one place.
SDK Updates
Custom Tools: Developers define functions that are exposed through a built‑in MCP server (custom-user-tools). These tools are automatically visible to all sub‑agents.
Auto‑Review Classifier: Natural‑language allow/block instructions placed in permissions.json steer the agent’s automatic review decisions.
Subagent Nesting: No depth limit (unlimited nesting).
Aider
No architectural changes. Model‑support updates include GPT‑5 family, Grok‑4, and Kimi K2.
Gap Analysis (vs. earlier implementations)
Based on a comparison with an experimental agent coding assistant (Day 126 work), the following gaps were identified in current tools (now addressed or still missing):
Feature
Status
Shell mode output fed back into conversation on demand
✅ New in Claude Code
Event‑triggered automations
✅ New in Cursor v3.8
Portable plugin/bundle format
✅ New in Cursor v3.9
Natural‑language‑steered permission classifier
✅ New in Cursor SDK
/rewind resume capability
✅ Parity achieved
! passthrough
✅ Parity achieved
Worktree spawn handoff
✅ Parity achieved
Draft‑PR handoff
✅ Parity achieved
The remaining gaps (from the perspective of the earlier experimental work) were filled by the June–July 2026 releases, bringing industry tools to parity or beyond.